chore(main): release 0.22.2 (#756)

* multitrack mixdown optimisation

* skip mixdown for multitrack

* skip mixdown for multitrack

---------

Co-authored-by: Igor Loskutov <igor.loskutoff@gmail.com>

.github/workflows/db_migrations.yml

@@ -2,6 +2,8 @@ name: Test Database Migrations
 
 on:
   push:
+    branches:
+      - main
     paths:
       - "server/migrations/**"
       - "server/reflector/db/**"
@@ -17,6 +19,9 @@ on:
 jobs:
   test-migrations:
     runs-on: ubuntu-latest
+    concurrency:
+      group: db-ubuntu-latest-${{ github.ref }}
+      cancel-in-progress: true
     services:
       postgres:
         image: postgres:17

.github/workflows/deploy.yml

@@ -1,4 +1,4 @@
-name: Deploy to Amazon ECS
+name: Build container/push to container registry
 
 on: [workflow_dispatch]
 

.github/workflows/docker-frontend.yml

@@ -0,0 +1,57 @@
+name: Build and Push Frontend Docker Image
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'www/**'
+      - '.github/workflows/docker-frontend.yml'
+  workflow_dispatch:
+
+env:
+  REGISTRY: ghcr.io
+  IMAGE_NAME: ${{ github.repository }}-frontend
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      packages: write
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Log in to GitHub Container Registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Extract metadata
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}
+          tags: |
+            type=ref,event=branch
+            type=sha,prefix={{branch}}-
+            type=raw,value=latest,enable={{is_default_branch}}
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build and push Docker image
+        uses: docker/build-push-action@v5
+        with:
+          context: ./www
+          file: ./www/Dockerfile
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          cache-from: type=gha
+          cache-to: type=gha,mode=max
+          platforms: linux/amd64,linux/arm64

.github/workflows/test_next_server.yml

@@ -0,0 +1,45 @@
+name: Test Next Server
+
+on:
+  pull_request:
+    paths:
+      - "www/**"
+  push:
+    branches:
+      - main
+    paths:
+      - "www/**"
+
+jobs:
+  test-next-server:
+    runs-on: ubuntu-latest
+
+    defaults:
+      run:
+        working-directory: ./www
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Setup Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+
+    - name: Install pnpm
+      uses: pnpm/action-setup@v4
+      with:
+        version: 8
+
+    - name: Setup Node.js cache
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'pnpm'
+        cache-dependency-path: './www/pnpm-lock.yaml'
+
+    - name: Install dependencies
+      run: pnpm install
+
+    - name: Run tests
+      run: pnpm test

.github/workflows/test_server.yml

@@ -5,12 +5,17 @@ on:
     paths:
       - "server/**"
   push:
+    branches:
+      - main
     paths:
       - "server/**"
 
 jobs:
   pytest:
     runs-on: ubuntu-latest
+    concurrency:
+      group: pytest-${{ github.ref }}
+      cancel-in-progress: true
     services:
       redis:
         image: redis:6
@@ -30,6 +35,9 @@ jobs:
 
   docker-amd64:
     runs-on: linux-amd64
+    concurrency:
+      group: docker-amd64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
       - name: Set up Docker Buildx
@@ -45,6 +53,9 @@ jobs:
 
   docker-arm64:
     runs-on: linux-arm64
+    concurrency:
+      group: docker-arm64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
       - name: Set up Docker Buildx

.gitignore

@@ -14,4 +14,7 @@ data/
 www/REFACTOR.md
 www/reload-frontend
 server/test.sqlite
-CLAUDE.local.md
+CLAUDE.local.md
+www/.env.development
+www/.env.production
+.playwright-mcp

.gitleaksignore

@@ -0,0 +1 @@
+b9d891d3424f371642cb032ecfd0e2564470a72c:server/tests/test_transcripts_recording_deletion.py:generic-api-key:15

.pre-commit-config.yaml

@@ -27,3 +27,8 @@ repos:
         files: ^server/
       - id: ruff-format
         files: ^server/
+
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.28.0
+    hooks:
+      - id: gitleaks

CHANGELOG.md

@@ -1,5 +1,239 @@
 # Changelog
 
+## [0.22.2](https://github.com/Monadical-SAS/reflector/compare/v0.22.1...v0.22.2) (2025-12-02)
+
+
+### Bug Fixes
+
+* daily auto refresh fix ([#755](https://github.com/Monadical-SAS/reflector/issues/755)) ([fe47c46](https://github.com/Monadical-SAS/reflector/commit/fe47c46489c5aa0cc538109f7559cc9accb35c01))
+* Skip mixdown for multitrack ([#760](https://github.com/Monadical-SAS/reflector/issues/760)) ([b51b7aa](https://github.com/Monadical-SAS/reflector/commit/b51b7aa9176c1a53ba57ad99f5e976c804a1e80c))
+
+## [0.22.1](https://github.com/Monadical-SAS/reflector/compare/v0.22.0...v0.22.1) (2025-11-27)
+
+
+### Bug Fixes
+
+* participants update from daily ([#749](https://github.com/Monadical-SAS/reflector/issues/749)) ([7f0b728](https://github.com/Monadical-SAS/reflector/commit/7f0b728991c1b9f9aae702c96297eae63b561ef5))
+
+## [0.22.0](https://github.com/Monadical-SAS/reflector/compare/v0.21.0...v0.22.0) (2025-11-26)
+
+
+### Features
+
+* Multitrack segmentation ([#747](https://github.com/Monadical-SAS/reflector/issues/747)) ([d63040e](https://github.com/Monadical-SAS/reflector/commit/d63040e2fdc07e7b272e85a39eb2411cd6a14798))
+
+## [0.21.0](https://github.com/Monadical-SAS/reflector/compare/v0.20.0...v0.21.0) (2025-11-26)
+
+
+### Features
+
+* add transcript format parameter to GET endpoint ([#709](https://github.com/Monadical-SAS/reflector/issues/709)) ([f6ca075](https://github.com/Monadical-SAS/reflector/commit/f6ca07505f34483b02270a2ef3bd809e9d2e1045))
+
+## [0.20.0](https://github.com/Monadical-SAS/reflector/compare/v0.19.0...v0.20.0) (2025-11-25)
+
+
+### Features
+
+* link transcript participants ([#737](https://github.com/Monadical-SAS/reflector/issues/737)) ([9bec398](https://github.com/Monadical-SAS/reflector/commit/9bec39808fc6322612d8b87e922a6f7901fc01c1))
+* transcript restart script ([#742](https://github.com/Monadical-SAS/reflector/issues/742)) ([86d5e26](https://github.com/Monadical-SAS/reflector/commit/86d5e26224bb55a0f1cc785aeda52065bb92ee6f))
+
+## [0.19.0](https://github.com/Monadical-SAS/reflector/compare/v0.18.0...v0.19.0) (2025-11-25)
+
+
+### Features
+
+* dailyco api module ([#725](https://github.com/Monadical-SAS/reflector/issues/725)) ([4287f8b](https://github.com/Monadical-SAS/reflector/commit/4287f8b8aeee60e51db7539f4dcbda5f6e696bd8))
+* dailyco poll ([#730](https://github.com/Monadical-SAS/reflector/issues/730)) ([8e438ca](https://github.com/Monadical-SAS/reflector/commit/8e438ca285152bd48fdc42767e706fb448d3525c))
+* multitrack cli ([#735](https://github.com/Monadical-SAS/reflector/issues/735)) ([11731c9](https://github.com/Monadical-SAS/reflector/commit/11731c9d38439b04e93b1c3afbd7090bad11a11f))
+
+
+### Bug Fixes
+
+* default platform fix ([#736](https://github.com/Monadical-SAS/reflector/issues/736)) ([c442a62](https://github.com/Monadical-SAS/reflector/commit/c442a627873ca667656eeaefb63e54ab10b8d19e))
+* parakeet vad not getting the end timestamp ([#728](https://github.com/Monadical-SAS/reflector/issues/728)) ([18ed713](https://github.com/Monadical-SAS/reflector/commit/18ed7133693653ef4ddac6c659a8c14b320d1657))
+* start raw tracks recording ([#729](https://github.com/Monadical-SAS/reflector/issues/729)) ([3e47c2c](https://github.com/Monadical-SAS/reflector/commit/3e47c2c0573504858e0d2e1798b6ed31f16b4a5d))
+
+## [0.18.0](https://github.com/Monadical-SAS/reflector/compare/v0.17.0...v0.18.0) (2025-11-14)
+
+
+### Features
+
+* daily QOL: participants dictionary ([#721](https://github.com/Monadical-SAS/reflector/issues/721)) ([b20cad7](https://github.com/Monadical-SAS/reflector/commit/b20cad76e69fb6a76405af299a005f1ddcf60eae))
+
+
+### Bug Fixes
+
+* add proccessing page to file upload and reprocessing ([#650](https://github.com/Monadical-SAS/reflector/issues/650)) ([28a7258](https://github.com/Monadical-SAS/reflector/commit/28a7258e45317b78e60e6397be2bc503647eaace))
+* copy transcript ([#674](https://github.com/Monadical-SAS/reflector/issues/674)) ([a9a4f32](https://github.com/Monadical-SAS/reflector/commit/a9a4f32324f66c838e081eee42bb9502f38c1db1))
+
+## [0.17.0](https://github.com/Monadical-SAS/reflector/compare/v0.16.0...v0.17.0) (2025-11-13)
+
+
+### Features
+
+* add API key management UI ([#716](https://github.com/Monadical-SAS/reflector/issues/716)) ([372202b](https://github.com/Monadical-SAS/reflector/commit/372202b0e1a86823900b0aa77be1bfbc2893d8a1))
+* daily.co support as alternative to whereby ([#691](https://github.com/Monadical-SAS/reflector/issues/691)) ([1473fd8](https://github.com/Monadical-SAS/reflector/commit/1473fd82dc472c394cbaa2987212ad662a74bcac))
+
+## [0.16.0](https://github.com/Monadical-SAS/reflector/compare/v0.15.0...v0.16.0) (2025-10-24)
+
+
+### Features
+
+* search date filter ([#710](https://github.com/Monadical-SAS/reflector/issues/710)) ([962c40e](https://github.com/Monadical-SAS/reflector/commit/962c40e2b6428ac42fd10aea926782d7a6f3f902))
+
+## [0.15.0](https://github.com/Monadical-SAS/reflector/compare/v0.14.0...v0.15.0) (2025-10-20)
+
+
+### Features
+
+* api tokens  ([#705](https://github.com/Monadical-SAS/reflector/issues/705)) ([9a258ab](https://github.com/Monadical-SAS/reflector/commit/9a258abc0209b0ac3799532a507ea6a9125d703a))
+
+## [0.14.0](https://github.com/Monadical-SAS/reflector/compare/v0.13.1...v0.14.0) (2025-10-08)
+
+
+### Features
+
+* Add calendar event data to transcript webhook payload ([#689](https://github.com/Monadical-SAS/reflector/issues/689)) ([5f6910e](https://github.com/Monadical-SAS/reflector/commit/5f6910e5131b7f28f86c9ecdcc57fed8412ee3cd))
+* container build for www / github ([#672](https://github.com/Monadical-SAS/reflector/issues/672)) ([969bd84](https://github.com/Monadical-SAS/reflector/commit/969bd84fcc14851d1a101412a0ba115f1b7cde82))
+* docker-compose for production frontend ([#664](https://github.com/Monadical-SAS/reflector/issues/664)) ([5bf64b5](https://github.com/Monadical-SAS/reflector/commit/5bf64b5a41f64535e22849b4bb11734d4dbb4aae))
+
+
+### Bug Fixes
+
+* restore feature boolean logic ([#671](https://github.com/Monadical-SAS/reflector/issues/671)) ([3660884](https://github.com/Monadical-SAS/reflector/commit/36608849ec64e953e3be456172502762e3c33df9))
+* security review ([#656](https://github.com/Monadical-SAS/reflector/issues/656)) ([5d98754](https://github.com/Monadical-SAS/reflector/commit/5d98754305c6c540dd194dda268544f6d88bfaf8))
+* update transcript list on reprocess ([#676](https://github.com/Monadical-SAS/reflector/issues/676)) ([9a71af1](https://github.com/Monadical-SAS/reflector/commit/9a71af145ee9b833078c78d0c684590ab12e9f0e))
+* upgrade nemo toolkit ([#678](https://github.com/Monadical-SAS/reflector/issues/678)) ([eef6dc3](https://github.com/Monadical-SAS/reflector/commit/eef6dc39037329b65804297786d852dddb0557f9))
+
+## [0.13.1](https://github.com/Monadical-SAS/reflector/compare/v0.13.0...v0.13.1) (2025-09-22)
+
+
+### Bug Fixes
+
+* TypeError on not all arguments converted during string formatting in logger ([#667](https://github.com/Monadical-SAS/reflector/issues/667)) ([565a629](https://github.com/Monadical-SAS/reflector/commit/565a62900f5a02fc946b68f9269a42190ed70ab6))
+
+## [0.13.0](https://github.com/Monadical-SAS/reflector/compare/v0.12.1...v0.13.0) (2025-09-19)
+
+
+### Features
+
+* room form edit with enter ([#662](https://github.com/Monadical-SAS/reflector/issues/662)) ([47716f6](https://github.com/Monadical-SAS/reflector/commit/47716f6e5ddee952609d2fa0ffabdfa865286796))
+
+
+### Bug Fixes
+
+* invalid cleanup call ([#660](https://github.com/Monadical-SAS/reflector/issues/660)) ([0abcebf](https://github.com/Monadical-SAS/reflector/commit/0abcebfc9491f87f605f21faa3e53996fafedd9a))
+
+## [0.12.1](https://github.com/Monadical-SAS/reflector/compare/v0.12.0...v0.12.1) (2025-09-17)
+
+
+### Bug Fixes
+
+* production blocked because having existing meeting with room_id null ([#657](https://github.com/Monadical-SAS/reflector/issues/657)) ([870e860](https://github.com/Monadical-SAS/reflector/commit/870e8605171a27155a9cbee215eeccb9a8d6c0a2))
+
+## [0.12.0](https://github.com/Monadical-SAS/reflector/compare/v0.11.0...v0.12.0) (2025-09-17)
+
+
+### Features
+
+* calendar integration ([#608](https://github.com/Monadical-SAS/reflector/issues/608)) ([6f680b5](https://github.com/Monadical-SAS/reflector/commit/6f680b57954c688882c4ed49f40f161c52a00a24))
+* self-hosted gpu api ([#636](https://github.com/Monadical-SAS/reflector/issues/636)) ([ab859d6](https://github.com/Monadical-SAS/reflector/commit/ab859d65a6bded904133a163a081a651b3938d42))
+
+
+### Bug Fixes
+
+* ignore player hotkeys for text inputs ([#646](https://github.com/Monadical-SAS/reflector/issues/646)) ([fa049e8](https://github.com/Monadical-SAS/reflector/commit/fa049e8d068190ce7ea015fd9fcccb8543f54a3f))
+
+## [0.11.0](https://github.com/Monadical-SAS/reflector/compare/v0.10.0...v0.11.0) (2025-09-16)
+
+
+### Features
+
+* remove profanity filter that was there for conference ([#652](https://github.com/Monadical-SAS/reflector/issues/652)) ([b42f7cf](https://github.com/Monadical-SAS/reflector/commit/b42f7cfc606783afcee792590efcc78b507468ab))
+
+
+### Bug Fixes
+
+* zulip and consent handler on the file pipeline ([#645](https://github.com/Monadical-SAS/reflector/issues/645)) ([5f143fe](https://github.com/Monadical-SAS/reflector/commit/5f143fe3640875dcb56c26694254a93189281d17))
+* zulip stream and topic selection in share dialog ([#644](https://github.com/Monadical-SAS/reflector/issues/644)) ([c546e69](https://github.com/Monadical-SAS/reflector/commit/c546e69739e68bb74fbc877eb62609928e5b8de6))
+
+## [0.10.0](https://github.com/Monadical-SAS/reflector/compare/v0.9.0...v0.10.0) (2025-09-11)
+
+
+### Features
+
+* replace nextjs-config with environment variables ([#632](https://github.com/Monadical-SAS/reflector/issues/632)) ([369ecdf](https://github.com/Monadical-SAS/reflector/commit/369ecdff13f3862d926a9c0b87df52c9d94c4dde))
+
+
+### Bug Fixes
+
+* anonymous users transcript permissions ([#621](https://github.com/Monadical-SAS/reflector/issues/621)) ([f81fe99](https://github.com/Monadical-SAS/reflector/commit/f81fe9948a9237b3e0001b2d8ca84f54d76878f9))
+* auth post ([#624](https://github.com/Monadical-SAS/reflector/issues/624)) ([cde99ca](https://github.com/Monadical-SAS/reflector/commit/cde99ca2716f84ba26798f289047732f0448742e))
+* auth post ([#626](https://github.com/Monadical-SAS/reflector/issues/626)) ([3b85ff3](https://github.com/Monadical-SAS/reflector/commit/3b85ff3bdf4fb053b103070646811bc990c0e70a))
+* auth post ([#627](https://github.com/Monadical-SAS/reflector/issues/627)) ([962038e](https://github.com/Monadical-SAS/reflector/commit/962038ee3f2a555dc3c03856be0e4409456e0996))
+* missing follow_redirects=True on modal endpoint ([#630](https://github.com/Monadical-SAS/reflector/issues/630)) ([fc363bd](https://github.com/Monadical-SAS/reflector/commit/fc363bd49b17b075e64f9186e5e0185abc325ea7))
+* sync backend and frontend token refresh logic ([#614](https://github.com/Monadical-SAS/reflector/issues/614)) ([5a5b323](https://github.com/Monadical-SAS/reflector/commit/5a5b3233820df9536da75e87ce6184a983d4713a))
+
+## [0.9.0](https://github.com/Monadical-SAS/reflector/compare/v0.8.2...v0.9.0) (2025-09-06)
+
+
+### Features
+
+* frontend openapi react query ([#606](https://github.com/Monadical-SAS/reflector/issues/606)) ([c4d2825](https://github.com/Monadical-SAS/reflector/commit/c4d2825c81f81ad8835629fbf6ea8c7383f8c31b))
+
+
+### Bug Fixes
+
+* align whisper transcriber api with parakeet ([#602](https://github.com/Monadical-SAS/reflector/issues/602)) ([0663700](https://github.com/Monadical-SAS/reflector/commit/0663700a615a4af69a03c96c410f049e23ec9443))
+* kv use tls explicit ([#610](https://github.com/Monadical-SAS/reflector/issues/610)) ([08d88ec](https://github.com/Monadical-SAS/reflector/commit/08d88ec349f38b0d13e0fa4cb73486c8dfd31836))
+* source kind for file processing ([#601](https://github.com/Monadical-SAS/reflector/issues/601)) ([dc82f8b](https://github.com/Monadical-SAS/reflector/commit/dc82f8bb3bdf3ab3d4088e592a30fd63907319e1))
+* token refresh locking ([#613](https://github.com/Monadical-SAS/reflector/issues/613)) ([7f5a4c9](https://github.com/Monadical-SAS/reflector/commit/7f5a4c9ddc7fd098860c8bdda2ca3b57f63ded2f))
+
+## [0.8.2](https://github.com/Monadical-SAS/reflector/compare/v0.8.1...v0.8.2) (2025-08-29)
+
+
+### Bug Fixes
+
+* search-logspam ([#593](https://github.com/Monadical-SAS/reflector/issues/593)) ([695d1a9](https://github.com/Monadical-SAS/reflector/commit/695d1a957d4cd862753049f9beed88836cabd5ab))
+
+## [0.8.1](https://github.com/Monadical-SAS/reflector/compare/v0.8.0...v0.8.1) (2025-08-29)
+
+
+### Bug Fixes
+
+* make webhook secret/url allowing null ([#590](https://github.com/Monadical-SAS/reflector/issues/590)) ([84a3812](https://github.com/Monadical-SAS/reflector/commit/84a381220bc606231d08d6f71d4babc818fa3c75))
+
+## [0.8.0](https://github.com/Monadical-SAS/reflector/compare/v0.7.3...v0.8.0) (2025-08-29)
+
+
+### Features
+
+* **cleanup:** add automatic data retention for public instances ([#574](https://github.com/Monadical-SAS/reflector/issues/574)) ([6f0c7c1](https://github.com/Monadical-SAS/reflector/commit/6f0c7c1a5e751713366886c8e764c2009e12ba72))
+* **rooms:** add webhook for transcript completion ([#578](https://github.com/Monadical-SAS/reflector/issues/578)) ([88ed7cf](https://github.com/Monadical-SAS/reflector/commit/88ed7cfa7804794b9b54cad4c3facc8a98cf85fd))
+
+
+### Bug Fixes
+
+* file pipeline status reporting and websocket updates ([#589](https://github.com/Monadical-SAS/reflector/issues/589)) ([9dfd769](https://github.com/Monadical-SAS/reflector/commit/9dfd76996f851cc52be54feea078adbc0816dc57))
+* Igor/evaluation ([#575](https://github.com/Monadical-SAS/reflector/issues/575)) ([124ce03](https://github.com/Monadical-SAS/reflector/commit/124ce03bf86044c18313d27228a25da4bc20c9c5))
+* optimize parakeet transcription batching algorithm ([#577](https://github.com/Monadical-SAS/reflector/issues/577)) ([7030e0f](https://github.com/Monadical-SAS/reflector/commit/7030e0f23649a8cf6c1eb6d5889684a41ce849ec))
+
+## [0.7.3](https://github.com/Monadical-SAS/reflector/compare/v0.7.2...v0.7.3) (2025-08-22)
+
+
+### Bug Fixes
+
+* cleaned repo, and get git-leaks clean ([359280d](https://github.com/Monadical-SAS/reflector/commit/359280dd340433ba4402ed69034094884c825e67))
+* restore previous behavior on live pipeline + audio downscaler ([#561](https://github.com/Monadical-SAS/reflector/issues/561)) ([9265d20](https://github.com/Monadical-SAS/reflector/commit/9265d201b590d23c628c5f19251b70f473859043))
+
+## [0.7.2](https://github.com/Monadical-SAS/reflector/compare/v0.7.1...v0.7.2) (2025-08-21)
+
+
+### Bug Fixes
+
+* docker image not loading libgomp.so.1 for torch ([#560](https://github.com/Monadical-SAS/reflector/issues/560)) ([773fccd](https://github.com/Monadical-SAS/reflector/commit/773fccd93e887c3493abc2e4a4864dddce610177))
+* include shared rooms to search ([#558](https://github.com/Monadical-SAS/reflector/issues/558)) ([499eced](https://github.com/Monadical-SAS/reflector/commit/499eced3360b84fb3a90e1c8a3b554290d21adc2))
+
 ## [0.7.1](https://github.com/Monadical-SAS/reflector/compare/v0.7.0...v0.7.1) (2025-08-21)
 
 

CLAUDE.md

@@ -66,7 +66,6 @@ pnpm install
 
 # Copy configuration templates
 cp .env_template .env
-cp config-template.ts config.ts
 ```
 
 **Development:**
@@ -152,7 +151,7 @@ All endpoints prefixed `/v1/`:
 
 **Frontend** (`www/.env`):
 - `NEXTAUTH_URL`, `NEXTAUTH_SECRET` - Authentication configuration
-- `NEXT_PUBLIC_REFLECTOR_API_URL` - Backend API endpoint
+- `REFLECTOR_API_URL` - Backend API endpoint
 - `REFLECTOR_DOMAIN_CONFIG` - Feature flags and domain settings
 
 ## Testing Strategy

README.md

@@ -1,43 +1,60 @@
 <div align="center">
+<img width="100" alt="image" src="https://github.com/user-attachments/assets/66fb367b-2c89-4516-9912-f47ac59c6a7f"/>
 
 # Reflector
 
-Reflector Audio Management and Analysis is a cutting-edge web application under development by Monadical. It utilizes AI to record meetings, providing a permanent record with transcripts, translations, and automated summaries.
+Reflector is an AI-powered audio transcription and meeting analysis platform that provides real-time transcription, speaker diarization, translation and summarization for audio content and live meetings. It works 100% with local models (whisper/parakeet, pyannote, seamless-m4t, and your local llm like phi-4).
 
-[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml)
+[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](https://opensource.org/licenses/MIT)
 </div>
-
-## Screenshots
+</div>
 <table>
   <tr>
     <td>
-      <a href="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3" />
+      <a href="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33" />
+      <a href="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc" />
+      <a href="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897" />
+      </a>
+    </td>
+    <td>
+      <a href="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4" />
       </a>
     </td>
   </tr>
 </table>
 
+## What is Reflector?
+
+Reflector is a web application that utilizes local models to process audio content, providing:
+
+- **Real-time Transcription**: Convert speech to text using [Whisper](https://github.com/openai/whisper) (multi-language) or [Parakeet](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) (English) models
+- **Speaker Diarization**: Identify and label different speakers using [Pyannote](https://github.com/pyannote/pyannote-audio) 3.1
+- **Live Translation**: Translate audio content in real-time to many languages with [Facebook Seamless-M4T](https://github.com/facebookresearch/seamless_communication)
+- **Topic Detection & Summarization**: Extract key topics and generate concise summaries using LLMs
+- **Meeting Recording**: Create permanent records of meetings with searchable transcripts
+
+Currently we provide [modal.com](https://modal.com/) gpu template to deploy.
+
 ## Background
 
 The project architecture consists of three primary components:
 
-- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
 - **Back-End**: Python server that offers an API and data persistence, found in `server/`.
-- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations. Most reliable option is Modal deployment
+- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
+- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations.
 
-It also uses authentik for authentication if activated, and Vercel for deployment and configuration of the front-end.
+It also uses authentik for authentication if activated.
 
 ## Contribution Guidelines
 
@@ -72,6 +89,8 @@ Note: We currently do not have instructions for Windows users.
 
 ## Installation
 
+*Note: we're working toward better installation, theses instructions are not accurate for now*
+
 ### Frontend
 
 Start with `cd www`.
@@ -80,11 +99,10 @@ Start with `cd www`.
 
 ```bash
 pnpm install
-cp .env_template .env
-cp config-template.ts config.ts
+cp .env.example .env
 ```
 
-Then, fill in the environment variables in `.env` and the configuration in `config.ts` as needed. If you are unsure on how to proceed, ask in Zulip.
+Then, fill in the environment variables in `.env` as needed. If you are unsure on how to proceed, ask in Zulip.
 
 **Run in development mode**
 
@@ -149,3 +167,47 @@ You can manually process an audio file by calling the process tool:
 ```bash
 uv run python -m reflector.tools.process path/to/audio.wav
 ```
+
+## Reprocessing any transcription
+
+```bash
+uv run -m reflector.tools.process_transcript 81ec38d1-9dd7-43d2-b3f8-51f4d34a07cd --sync
+```
+
+## Build-time env variables
+
+Next.js projects are more used to NEXT_PUBLIC_ prefixed buildtime vars. We don't have those for the reason we need to serve a ccustomizable prebuild docker container.
+
+Instead, all the variables are runtime. Variables needed to the frontend are served to the frontend app at initial render.
+
+It also means there's no static prebuild and no static files to serve for js/html.
+
+## Feature Flags
+
+Reflector uses environment variable-based feature flags to control application functionality. These flags allow you to enable or disable features without code changes.
+
+### Available Feature Flags
+
+| Feature Flag | Environment Variable |
+|-------------|---------------------|
+| `requireLogin` | `FEATURE_REQUIRE_LOGIN` |
+| `privacy` | `FEATURE_PRIVACY` |
+| `browse` | `FEATURE_BROWSE` |
+| `sendToZulip` | `FEATURE_SEND_TO_ZULIP` |
+| `rooms` | `FEATURE_ROOMS` |
+
+### Setting Feature Flags
+
+Feature flags are controlled via environment variables using the pattern `FEATURE_{FEATURE_NAME}` where `{FEATURE_NAME}` is the SCREAMING_SNAKE_CASE version of the feature name.
+
+**Examples:**
+```bash
+# Enable user authentication requirement
+FEATURE_REQUIRE_LOGIN=true
+
+# Disable browse functionality
+FEATURE_BROWSE=false
+
+# Enable Zulip integration
+FEATURE_SEND_TO_ZULIP=true
+```

docker-compose.prod.yml

@@ -0,0 +1,39 @@
+# Production Docker Compose configuration for Frontend
+# Usage: docker compose -f docker-compose.prod.yml up -d
+
+services:
+  web:
+    build:
+      context: ./www
+      dockerfile: Dockerfile
+    image: reflector-frontend:latest
+    environment:
+      - KV_URL=${KV_URL:-redis://redis:6379}
+      - SITE_URL=${SITE_URL}
+      - API_URL=${API_URL}
+      - WEBSOCKET_URL=${WEBSOCKET_URL}
+      - NEXTAUTH_URL=${NEXTAUTH_URL:-http://localhost:3000}
+      - NEXTAUTH_SECRET=${NEXTAUTH_SECRET:-changeme-in-production}
+      - AUTHENTIK_ISSUER=${AUTHENTIK_ISSUER}
+      - AUTHENTIK_CLIENT_ID=${AUTHENTIK_CLIENT_ID}
+      - AUTHENTIK_CLIENT_SECRET=${AUTHENTIK_CLIENT_SECRET}
+      - AUTHENTIK_REFRESH_TOKEN_URL=${AUTHENTIK_REFRESH_TOKEN_URL}
+      - SENTRY_DSN=${SENTRY_DSN}
+      - SENTRY_IGNORE_API_RESOLUTION_ERROR=${SENTRY_IGNORE_API_RESOLUTION_ERROR:-1}
+    depends_on:
+      - redis
+    restart: unless-stopped
+
+  redis:
+    image: redis:7.2-alpine
+    restart: unless-stopped
+    healthcheck:
+      test: ["CMD", "redis-cli", "ping"]
+      interval: 30s
+      timeout: 3s
+      retries: 3
+    volumes:
+      - redis_data:/data
+
+volumes:
+  redis_data:

docker-compose.yml

@@ -6,6 +6,7 @@ services:
       - 1250:1250
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -16,6 +17,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -26,6 +28,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -36,7 +39,7 @@ services:
     ports:
       - 6379:6379
   web:
-    image: node:18
+    image: node:22-alpine
     ports:
       - "3000:3000"
     command: sh -c "corepack enable && pnpm install && pnpm dev"
@@ -47,6 +50,8 @@ services:
       - /app/node_modules
     env_file:
       - ./www/.env.local
+    environment:
+      - NODE_ENV=development
 
   postgres:
     image: postgres:17

docs/transcript.md

@@ -0,0 +1,241 @@
+# Transcript Formats
+
+The Reflector API provides multiple output formats for transcript data through the `transcript_format` query parameter on the GET `/v1/transcripts/{id}` endpoint.
+
+## Overview
+
+When retrieving a transcript, you can specify the desired format using the `transcript_format` query parameter. The API supports four formats optimized for different use cases:
+
+- **text** - Plain text with speaker names (default)
+- **text-timestamped** - Timestamped text with speaker names
+- **webvtt-named** - WebVTT subtitle format with participant names
+- **json** - Structured JSON segments with full metadata
+
+All formats include participant information when available, resolving speaker IDs to actual names.
+
+## Query Parameter Usage
+
+```
+GET /v1/transcripts/{id}?transcript_format={format}
+```
+
+### Parameters
+
+- `transcript_format` (optional): The desired output format
+  - Type: `"text" | "text-timestamped" | "webvtt-named" | "json"`
+  - Default: `"text"`
+
+## Format Descriptions
+
+### Text Format (`text`)
+
+**Use case:** Simple, human-readable transcript for display or export.
+
+**Format:** Speaker names followed by their dialogue, one line per segment.
+
+**Example:**
+```
+John Smith: Hello everyone
+Jane Doe: Hi there
+John Smith: How are you today?
+```
+
+**Request:**
+```bash
+GET /v1/transcripts/{id}?transcript_format=text
+```
+
+**Response:**
+```json
+{
+  "id": "transcript_123",
+  "name": "Meeting Recording",
+  "transcript_format": "text",
+  "transcript": "John Smith: Hello everyone\nJane Doe: Hi there\nJohn Smith: How are you today?",
+  "participants": [
+    {"id": "p1", "speaker": 0, "name": "John Smith"},
+    {"id": "p2", "speaker": 1, "name": "Jane Doe"}
+  ],
+  ...
+}
+```
+
+### Text Timestamped Format (`text-timestamped`)
+
+**Use case:** Transcript with timing information for navigation or reference.
+
+**Format:** `[MM:SS]` timestamp prefix before each speaker and dialogue.
+
+**Example:**
+```
+[00:00] John Smith: Hello everyone
+[00:05] Jane Doe: Hi there
+[00:12] John Smith: How are you today?
+```
+
+**Request:**
+```bash
+GET /v1/transcripts/{id}?transcript_format=text-timestamped
+```
+
+**Response:**
+```json
+{
+  "id": "transcript_123",
+  "name": "Meeting Recording",
+  "transcript_format": "text-timestamped",
+  "transcript": "[00:00] John Smith: Hello everyone\n[00:05] Jane Doe: Hi there\n[00:12] John Smith: How are you today?",
+  "participants": [
+    {"id": "p1", "speaker": 0, "name": "John Smith"},
+    {"id": "p2", "speaker": 1, "name": "Jane Doe"}
+  ],
+  ...
+}
+```
+
+### WebVTT Named Format (`webvtt-named`)
+
+**Use case:** Subtitle files for video players, accessibility tools, or video editing.
+
+**Format:** Standard WebVTT subtitle format with voice tags using participant names.
+
+**Example:**
+```
+WEBVTT
+
+00:00:00.000 --> 00:00:05.000
+<v John Smith>Hello everyone
+
+00:00:05.000 --> 00:00:12.000
+<v Jane Doe>Hi there
+
+00:00:12.000 --> 00:00:18.000
+<v John Smith>How are you today?
+```
+
+**Request:**
+```bash
+GET /v1/transcripts/{id}?transcript_format=webvtt-named
+```
+
+**Response:**
+```json
+{
+  "id": "transcript_123",
+  "name": "Meeting Recording",
+  "transcript_format": "webvtt-named",
+  "transcript": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v John Smith>Hello everyone\n\n...",
+  "participants": [
+    {"id": "p1", "speaker": 0, "name": "John Smith"},
+    {"id": "p2", "speaker": 1, "name": "Jane Doe"}
+  ],
+  ...
+}
+```
+
+### JSON Format (`json`)
+
+**Use case:** Programmatic access with full timing and speaker metadata.
+
+**Format:** Array of segment objects with speaker information, text content, and precise timing.
+
+**Example:**
+```json
+[
+  {
+    "speaker": 0,
+    "speaker_name": "John Smith",
+    "text": "Hello everyone",
+    "start": 0.0,
+    "end": 5.0
+  },
+  {
+    "speaker": 1,
+    "speaker_name": "Jane Doe",
+    "text": "Hi there",
+    "start": 5.0,
+    "end": 12.0
+  },
+  {
+    "speaker": 0,
+    "speaker_name": "John Smith",
+    "text": "How are you today?",
+    "start": 12.0,
+    "end": 18.0
+  }
+]
+```
+
+**Request:**
+```bash
+GET /v1/transcripts/{id}?transcript_format=json
+```
+
+**Response:**
+```json
+{
+  "id": "transcript_123",
+  "name": "Meeting Recording",
+  "transcript_format": "json",
+  "transcript": [
+    {
+      "speaker": 0,
+      "speaker_name": "John Smith",
+      "text": "Hello everyone",
+      "start": 0.0,
+      "end": 5.0
+    },
+    {
+      "speaker": 1,
+      "speaker_name": "Jane Doe",
+      "text": "Hi there",
+      "start": 5.0,
+      "end": 12.0
+    }
+  ],
+  "participants": [
+    {"id": "p1", "speaker": 0, "name": "John Smith"},
+    {"id": "p2", "speaker": 1, "name": "Jane Doe"}
+  ],
+  ...
+}
+```
+
+## Response Structure
+
+All formats return the same base transcript metadata with an additional `transcript_format` field and format-specific `transcript` field:
+
+### Common Fields
+
+- `id`: Transcript identifier
+- `user_id`: Owner user ID (if authenticated)
+- `name`: Transcript name
+- `status`: Processing status
+- `locked`: Whether transcript is locked for editing
+- `duration`: Total duration in seconds
+- `title`: Auto-generated or custom title
+- `short_summary`: Brief summary
+- `long_summary`: Detailed summary
+- `created_at`: Creation timestamp
+- `share_mode`: Access control setting
+- `source_language`: Original audio language
+- `target_language`: Translation target language
+- `reviewed`: Whether transcript has been reviewed
+- `meeting_id`: Associated meeting ID (if applicable)
+- `source_kind`: Source type (live, file, room)
+- `room_id`: Associated room ID (if applicable)
+- `audio_deleted`: Whether audio has been deleted
+- `participants`: Array of participant objects with speaker mappings
+
+### Format-Specific Fields
+
+- `transcript_format`: The format identifier (discriminator field)
+- `transcript`: The formatted transcript content (string for text/webvtt formats, array for json format)
+
+## Speaker Name Resolution
+
+All formats resolve speaker IDs to participant names when available:
+
+- If a participant exists for the speaker ID, their name is used
+- If no participant exists, a default name like "Speaker 0" is generated
+- Speaker IDs are integers (0, 1, 2, etc.) assigned during diarization

gpu/modal_deployments/.gitignore

@@ -0,0 +1,33 @@
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Logs
+*.log
+
+# Env and secrets
+.env
+.env.*
+*.env
+*.secret
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Modal local state (if any)
+modal_mounts/
+.modal_cache/

gpu/modal_deployments/reflector_transcriber.py

@@ -0,0 +1,608 @@
+import os
+import sys
+import threading
+import uuid
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
+from urllib.parse import urlparse
+
+import modal
+
+MODEL_NAME = "large-v2"
+MODEL_COMPUTE_TYPE: str = "float16"
+MODEL_NUM_WORKERS: int = 1
+MINUTES = 60  # seconds
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/models"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+
+WhisperUniqFilename = NewType("WhisperUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App("reflector-transcriber")
+
+model_cache = modal.Volume.from_name("models", create_if_missing=True)
+upload_volume = modal.Volume.from_name("whisper-uploads", create_if_missing=True)
+
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
+def download_model():
+    from faster_whisper import download_model
+
+    model_cache.reload()
+
+    download_model(MODEL_NAME, cache_dir=CACHE_PATH)
+
+    model_cache.commit()
+
+
+image = (
+    modal.Image.debian_slim(python_version="3.12")
+    .env(
+        {
+            "HF_HUB_ENABLE_HF_TRANSFER": "1",
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
+            ),
+        }
+    )
+    .apt_install("ffmpeg")
+    .pip_install(
+        "huggingface_hub==0.27.1",
+        "hf-transfer==0.1.9",
+        "torch==2.5.1",
+        "faster-whisper==1.1.1",
+        "fastapi==0.115.12",
+        "requests",
+        "librosa==0.10.1",
+        "numpy<2",
+        "silero-vad==5.1.0",
+    )
+    .run_function(download_model, volumes={CACHE_PATH: model_cache})
+)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
+    parsed_url = urlparse(url)
+    url_path = parsed_url.path
+
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return AudioFileExtension(ext)
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return AudioFileExtension("mp3")
+    if "audio/wav" in content_type:
+        return AudioFileExtension("wav")
+    if "audio/mp4" in content_type:
+        return AudioFileExtension("mp4")
+
+    raise ValueError(
+        f"Unsupported audio format for URL: {url}. "
+        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+    )
+
+
+def download_audio_to_volume(
+    audio_file_url: str,
+) -> tuple[WhisperUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = WhisperUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
+    """Add 0.5s of silence if audio is shorter than the silence_padding window.
+
+    Whisper does not require this strictly, but aligning behavior with Parakeet
+    avoids edge-case crashes on extremely short inputs and makes comparisons easier.
+    """
+    import numpy as np
+
+    audio_duration = len(audio_array) / sample_rate
+    if audio_duration < VAD_CONFIG["silence_padding"]:
+        silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+        silence = np.zeros(silence_samples, dtype=np.float32)
+        return np.concatenate([audio_array, silence])
+    return audio_array
+
+
+@app.cls(
+    gpu="A10G",
+    timeout=5 * MINUTES,
+    scaledown_window=5 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+@modal.concurrent(max_inputs=10)
+class TranscriberWhisperLive:
+    """Live transcriber class for small audio segments (A10G).
+
+    Mirrors the Parakeet live class API but uses Faster-Whisper under the hood.
+    """
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        print(f"Model is on device: {self.device}")
+
+    @modal.method()
+    def transcribe_segment(
+        self,
+        filename: str,
+        language: str = "en",
+    ):
+        """Transcribe a single uploaded audio file by filename."""
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    file_path,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+
+        return {"text": text, "words": words}
+
+    @modal.method()
+    def transcribe_batch(
+        self,
+        filenames: list[str],
+        language: str = "en",
+    ):
+        """Transcribe multiple uploaded audio files and return per-file results."""
+        upload_volume.reload()
+
+        results = []
+        for filename in filenames:
+            file_path = f"{UPLOADS_PATH}/{filename}"
+            if not os.path.exists(file_path):
+                raise FileNotFoundError(f"Batch file not found: {file_path}")
+
+            with self.lock:
+                with NoStdStreams():
+                    segments, _ = self.model.transcribe(
+                        file_path,
+                        language=language,
+                        beam_size=5,
+                        word_timestamps=True,
+                        vad_filter=True,
+                        vad_parameters={"min_silence_duration_ms": 500},
+                    )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start), 2),
+                    "end": round(float(w.end), 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+
+            results.append(
+                {
+                    "filename": filename,
+                    "text": text,
+                    "words": words,
+                }
+            )
+
+        return results
+
+
+@app.cls(
+    gpu="L40S",
+    timeout=15 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+class TranscriberWhisperFile:
+    """File transcriber for larger/longer audio, using VAD-driven batching (L40S)."""
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+        from silero_vad import load_silero_vad
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        self.vad_model = load_silero_vad(onnx=False)
+
+    @modal.method()
+    def transcribe_segment(
+        self, filename: str, timestamp_offset: float = 0.0, language: str = "en"
+    ):
+        import librosa
+        import numpy as np
+        from silero_vad import VADIterator
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLERATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[TimeSegment, None, None]:
+            """Generate speech segments as TimeSegment using Silero VAD."""
+            iterator = VADIterator(self.vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield TimeSegment(
+                        start / float(SAMPLERATE), end / float(SAMPLERATE)
+                    )
+                    start = None
+            iterator.reset_states()
+
+        upload_volume.reload()
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        audio_array, _sr = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+
+        # Batch segments up to ~30s windows by merging contiguous VAD segments
+        merged_batches: list[TimeSegment] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for segment in vad_segments(audio_array):
+            seg_start, seg_end = segment.start, segment.end
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append(TimeSegment(batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append(TimeSegment(batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+
+        for segment in merged_batches:
+            start_time, end_time = segment.start, segment.end
+            s_idx = int(start_time * SAMPLERATE)
+            e_idx = int(end_time * SAMPLERATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = pad_audio(segment, SAMPLERATE)
+
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        return {"text": " ".join(all_text), "words": all_words}
+
+
+def detect_audio_format(url: str, headers: dict) -> str:
+    from urllib.parse import urlparse
+
+    from fastapi import HTTPException
+
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_volume(audio_file_url: str) -> tuple[str, str]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+@app.function(
+    scaledown_window=60,
+    timeout=600,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    image=image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import (
+        Body,
+        Depends,
+        FastAPI,
+        Form,
+        HTTPException,
+        UploadFile,
+        status,
+    )
+    from fastapi.security import OAuth2PasswordBearer
+
+    transcriber_live = TranscriberWhisperLive()
+    transcriber_file = TranscriberWhisperFile()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    class TranscriptResponse(dict):
+        pass
+
+    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
+    def transcribe(
+        file: UploadFile = None,
+        files: list[UploadFile] | None = None,
+        model: str = Form(MODEL_NAME),
+        language: str = Form("en"),
+        batch: bool = Form(False),
+    ):
+        if not file and not files:
+            raise HTTPException(
+                status_code=400, detail="Either 'file' or 'files' parameter is required"
+            )
+        if batch and not files:
+            raise HTTPException(
+                status_code=400, detail="Batch transcription requires 'files'"
+            )
+
+        upload_files = [file] if file else files
+
+        uploaded_filenames: list[str] = []
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1]
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = f"{UPLOADS_PATH}/{unique_filename}"
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_filenames.append(unique_filename)
+
+        upload_volume.commit()
+
+        try:
+            if batch and len(upload_files) > 1:
+                func = transcriber_live.transcribe_batch.spawn(
+                    filenames=uploaded_filenames,
+                    language=language,
+                )
+                results = func.get()
+                return {"results": results}
+
+            results = []
+            for filename in uploaded_filenames:
+                func = transcriber_live.transcribe_segment.spawn(
+                    filename=filename,
+                    language=language,
+                )
+                result = func.get()
+                result["filename"] = filename
+                results.append(result)
+
+            return {"results": results} if len(results) > 1 else results[0]
+        finally:
+            for filename in uploaded_filenames:
+                try:
+                    file_path = f"{UPLOADS_PATH}/{filename}"
+                    os.remove(file_path)
+                except Exception:
+                    pass
+            upload_volume.commit()
+
+    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
+    def transcribe_from_url(
+        audio_file_url: str = Body(
+            ..., description="URL of the audio file to transcribe"
+        ),
+        model: str = Body(MODEL_NAME),
+        language: str = Body("en"),
+        timestamp_offset: float = Body(0.0),
+    ):
+        unique_filename, _audio_suffix = download_audio_to_volume(audio_file_url)
+        try:
+            func = transcriber_file.transcribe_segment.spawn(
+                filename=unique_filename,
+                timestamp_offset=timestamp_offset,
+                language=language,
+            )
+            result = func.get()
+            return result
+        finally:
+            try:
+                file_path = f"{UPLOADS_PATH}/{unique_filename}"
+                os.remove(file_path)
+                upload_volume.commit()
+            except Exception:
+                pass
+
+    return app
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()

gpu/modal_deployments/reflector_transcriber_parakeet.py

@@ -3,7 +3,7 @@ import os
 import sys
 import threading
 import uuid
-from typing import Mapping, NewType
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
 from urllib.parse import urlparse
 
 import modal
@@ -14,10 +14,7 @@ SAMPLERATE = 16000
 UPLOADS_PATH = "/uploads"
 CACHE_PATH = "/cache"
 VAD_CONFIG = {
-    "max_segment_duration": 30.0,
-    "batch_max_files": 10,
-    "batch_max_duration": 5.0,
-    "min_segment_duration": 0.02,
+    "batch_max_duration": 30.0,
     "silence_padding": 0.5,
     "window_size": 512,
 }
@@ -25,6 +22,37 @@ VAD_CONFIG = {
 ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
 AudioFileExtension = NewType("AudioFileExtension", str)
 
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
 app = modal.App("reflector-transcriber-parakeet")
 
 # Volume for caching model weights
@@ -49,13 +77,13 @@ image = (
     .pip_install(
         "hf_transfer==0.1.9",
         "huggingface_hub[hf-xet]==0.31.2",
-        "nemo_toolkit[asr]==2.3.0",
+        "nemo_toolkit[asr]==2.5.0",
         "cuda-python==12.8.0",
         "fastapi==0.115.12",
         "numpy<2",
-        "librosa==0.10.1",
+        "librosa==0.11.0",
         "requests",
-        "silero-vad==5.1.0",
+        "silero-vad==6.2.0",
         "torch",
     )
     .entrypoint([])  # silence chatty logs by container on start
@@ -170,12 +198,14 @@ class TranscriberParakeetLive:
                 (output,) = self.model.transcribe([padded_audio], timestamps=True)
 
         text = output.text.strip()
-        words = [
-            {
-                "word": word_info["word"],
-                "start": round(word_info["start"], 2),
-                "end": round(word_info["end"], 2),
-            }
+        words: list[WordTiming] = [
+            WordTiming(
+                # XXX the space added here is to match the output of whisper
+                # whisper add space to each words, while parakeet don't
+                word=word_info["word"] + " ",
+                start=round(word_info["start"], 2),
+                end=round(word_info["end"], 2),
+            )
             for word_info in output.timestamp["word"]
         ]
 
@@ -211,12 +241,12 @@ class TranscriberParakeetLive:
         for i, (filename, output) in enumerate(zip(filenames, outputs)):
             text = output.text.strip()
 
-            words = [
-                {
-                    "word": word_info["word"],
-                    "start": round(word_info["start"], 2),
-                    "end": round(word_info["end"], 2),
-                }
+            words: list[WordTiming] = [
+                WordTiming(
+                    word=word_info["word"] + " ",
+                    start=round(word_info["start"], 2),
+                    end=round(word_info["end"], 2),
+                )
                 for word_info in output.timestamp["word"]
             ]
 
@@ -271,9 +301,12 @@ class TranscriberParakeetFile:
             audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
             return audio_array
 
-        def vad_segment_generator(audio_array):
+        def vad_segment_generator(
+            audio_array,
+        ) -> Generator[TimeSegment, None, None]:
             """Generate speech segments using VAD with start/end sample indices"""
             vad_iterator = VADIterator(self.vad_model, sampling_rate=SAMPLERATE)
+            audio_duration = len(audio_array) / float(SAMPLERATE)
             window_size = VAD_CONFIG["window_size"]
             start = None
 
@@ -297,107 +330,125 @@ class TranscriberParakeetFile:
                     start_time = start / float(SAMPLERATE)
                     end_time = end / float(SAMPLERATE)
 
-                    # Extract the actual audio segment
-                    audio_segment = audio_array[start:end]
-
-                    yield (start_time, end_time, audio_segment)
+                    yield TimeSegment(start_time, end_time)
                     start = None
 
+            if start is not None:
+                start_time = start / float(SAMPLERATE)
+                yield TimeSegment(start_time, audio_duration)
+
             vad_iterator.reset_states()
 
-        def vad_segment_filter(segments):
-            """Filter VAD segments by duration and chunk large segments"""
-            min_dur = VAD_CONFIG["min_segment_duration"]
-            max_dur = VAD_CONFIG["max_segment_duration"]
+        def batch_speech_segments(
+            segments: Generator[TimeSegment, None, None], max_duration: int
+        ) -> Generator[TimeSegment, None, None]:
+            """
+            Input segments:
+              [0-2] [3-5] [6-8] [10-11] [12-15] [17-19] [20-22]
 
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
+                                  ↓ (max_duration=10)
 
-                # Skip very small segments
-                if segment_duration < min_dur:
+              Output batches:
+              [0-8]           [10-19]          [20-22]
+
+            Note: silences are kept for better transcription, previous implementation was
+            passing segments separatly, but the output was less accurate.
+            """
+            batch_start_time = None
+            batch_end_time = None
+
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                if batch_start_time is None or batch_end_time is None:
+                    batch_start_time = start_time
+                    batch_end_time = end_time
                     continue
 
-                # If segment is within max duration, yield as-is
-                if segment_duration <= max_dur:
-                    yield (start_time, end_time, audio_segment)
+                total_duration = end_time - batch_start_time
+
+                if total_duration <= max_duration:
+                    batch_end_time = end_time
                     continue
 
-                # Chunk large segments into smaller pieces
-                chunk_samples = int(max_dur * SAMPLERATE)
-                current_start = start_time
+                yield TimeSegment(batch_start_time, batch_end_time)
+                batch_start_time = start_time
+                batch_end_time = end_time
 
-                for chunk_offset in range(0, len(audio_segment), chunk_samples):
-                    chunk_audio = audio_segment[
-                        chunk_offset : chunk_offset + chunk_samples
-                    ]
-                    if len(chunk_audio) == 0:
-                        break
+            if batch_start_time is None or batch_end_time is None:
+                return
 
-                    chunk_duration = len(chunk_audio) / float(SAMPLERATE)
-                    chunk_end = current_start + chunk_duration
+            yield TimeSegment(batch_start_time, batch_end_time)
 
-                    # Only yield chunks that meet minimum duration
-                    if chunk_duration >= min_dur:
-                        yield (current_start, chunk_end, chunk_audio)
+        def batch_segment_to_audio_segment(
+            segments: Generator[TimeSegment, None, None],
+            audio_array,
+        ) -> Generator[AudioSegment, None, None]:
+            """Extract audio segments and apply padding for Parakeet compatibility.
 
-                    current_start = chunk_end
+            Uses pad_audio to ensure segments are at least 0.5s long, preventing
+            Parakeet crashes. This padding may cause slight timing overlaps between
+            segments, which are corrected by enforce_word_timing_constraints.
+            """
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                start_sample = int(start_time * SAMPLERATE)
+                end_sample = int(end_time * SAMPLERATE)
+                audio_segment = audio_array[start_sample:end_sample]
 
-        def batch_segments(segments, max_files=10, max_duration=5.0):
-            batch = []
-            batch_duration = 0.0
+                padded_segment = pad_audio(audio_segment, SAMPLERATE)
 
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
+                yield AudioSegment(start_time, end_time, padded_segment)
 
-                if segment_duration < VAD_CONFIG["silence_padding"]:
-                    silence_samples = int(
-                        (VAD_CONFIG["silence_padding"] - segment_duration) * SAMPLERATE
-                    )
-                    padding = np.zeros(silence_samples, dtype=np.float32)
-                    audio_segment = np.concatenate([audio_segment, padding])
-                    segment_duration = VAD_CONFIG["silence_padding"]
-
-                batch.append((start_time, end_time, audio_segment))
-                batch_duration += segment_duration
-
-                if len(batch) >= max_files or batch_duration >= max_duration:
-                    yield batch
-                    batch = []
-                    batch_duration = 0.0
-
-            if batch:
-                yield batch
-
-        def transcribe_batch(model, audio_segments):
+        def transcribe_batch(model, audio_segments: list) -> list:
             with NoStdStreams():
                 outputs = model.transcribe(audio_segments, timestamps=True)
             return outputs
 
+        def enforce_word_timing_constraints(
+            words: list[WordTiming],
+        ) -> list[WordTiming]:
+            """Enforce that word end times don't exceed the start time of the next word.
+
+            Due to silence padding added in batch_segment_to_audio_segment for better
+            transcription accuracy, word timings from different segments may overlap.
+            This function ensures there are no overlaps by adjusting end times.
+            """
+            if len(words) <= 1:
+                return words
+
+            enforced_words = []
+            for i, word in enumerate(words):
+                enforced_word = word.copy()
+
+                if i < len(words) - 1:
+                    next_start = words[i + 1]["start"]
+                    if enforced_word["end"] > next_start:
+                        enforced_word["end"] = next_start
+
+                enforced_words.append(enforced_word)
+
+            return enforced_words
+
         def emit_results(
-            results,
-            segments_info,
-            batch_index,
-            total_batches,
-        ):
+            results: list,
+            segments_info: list[AudioSegment],
+        ) -> Generator[TranscriptResult, None, None]:
             """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
-            for i, (output, (start_time, end_time, _)) in enumerate(
-                zip(results, segments_info)
-            ):
+            for i, (output, segment) in enumerate(zip(results, segments_info)):
+                start_time, end_time = segment.start, segment.end
                 text = output.text.strip()
-                words = [
-                    {
-                        "word": word_info["word"],
-                        "start": round(
+                words: list[WordTiming] = [
+                    WordTiming(
+                        word=word_info["word"] + " ",
+                        start=round(
                             word_info["start"] + start_time + timestamp_offset, 2
                         ),
-                        "end": round(
-                            word_info["end"] + start_time + timestamp_offset, 2
-                        ),
-                    }
+                        end=round(word_info["end"] + start_time + timestamp_offset, 2),
+                    )
                     for word_info in output.timestamp["word"]
                 ]
 
-                yield text, words
+                yield TranscriptResult(text, words)
 
         upload_volume.reload()
 
@@ -407,41 +458,31 @@ class TranscriberParakeetFile:
 
         audio_array = load_and_convert_audio(file_path)
         total_duration = len(audio_array) / float(SAMPLERATE)
-        processed_duration = 0.0
 
-        all_text_parts = []
-        all_words = []
+        all_text_parts: list[str] = []
+        all_words: list[WordTiming] = []
 
         raw_segments = vad_segment_generator(audio_array)
-        filtered_segments = vad_segment_filter(raw_segments)
-        batches = batch_segments(
-            filtered_segments,
-            VAD_CONFIG["batch_max_files"],
+        speech_segments = batch_speech_segments(
+            raw_segments,
             VAD_CONFIG["batch_max_duration"],
         )
+        audio_segments = batch_segment_to_audio_segment(speech_segments, audio_array)
 
-        batch_index = 0
-        total_batches = max(
-            1, int(total_duration / VAD_CONFIG["batch_max_duration"]) + 1
-        )
+        for batch in audio_segments:
+            audio_segment = batch.audio
+            results = transcribe_batch(self.model, [audio_segment])
 
-        for batch in batches:
-            batch_index += 1
-            audio_segments = [seg[2] for seg in batch]
-            results = transcribe_batch(self.model, audio_segments)
-
-            for text, words in emit_results(
+            for result in emit_results(
                 results,
-                batch,
-                batch_index,
-                total_batches,
+                [batch],
             ):
-                if not text:
+                if not result.text:
                     continue
-                all_text_parts.append(text)
-                all_words.extend(words)
+                all_text_parts.append(result.text)
+                all_words.extend(result.words)
 
-            processed_duration += sum(len(seg[2]) / float(SAMPLERATE) for seg in batch)
+        all_words = enforce_word_timing_constraints(all_words)
 
         combined_text = " ".join(all_text_parts)
         return {"text": combined_text, "words": all_words}

gpu/self_hosted/.env.example

@@ -0,0 +1,2 @@
+REFLECTOR_GPU_APIKEY=
+HF_TOKEN=

gpu/self_hosted/.gitignore

@@ -0,0 +1,38 @@
+cache/
+
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Env and secrets
+.env
+*.env
+*.secret
+HF_TOKEN
+REFLECTOR_GPU_APIKEY
+
+# Virtual env / uv
+.venv/
+venv/
+ENV/
+uv/
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Logs
+*.log

gpu/self_hosted/Dockerfile

@@ -0,0 +1,46 @@
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 \
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
+
+WORKDIR /tmp
+RUN apt-get update \
+ && apt-get install -y \
+    ffmpeg \
+    curl \
+    ca-certificates \
+    gnupg \
+    wget \
+ && apt-get clean
+# Add NVIDIA CUDA repo for Debian 12 (bookworm) and install cuDNN 9 for CUDA 12
+ADD https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb /cuda-keyring.deb
+RUN dpkg -i /cuda-keyring.deb \
+ && rm /cuda-keyring.deb \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+    cuda-cudart-12-6 \
+    libcublas-12-6 \
+    libcudnn9-cuda-12 \
+    libcudnn9-dev-cuda-12 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+ADD https://astral.sh/uv/install.sh /uv-installer.sh
+RUN sh /uv-installer.sh && rm /uv-installer.sh
+ENV PATH="/root/.local/bin/:$PATH"
+ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH"
+
+RUN mkdir -p /app
+WORKDIR /app
+COPY pyproject.toml uv.lock /app/
+
+
+COPY ./app /app/app
+COPY ./main.py /app/
+COPY ./runserver.sh /app/
+
+EXPOSE 8000
+
+CMD ["sh", "/app/runserver.sh"]
+
+

gpu/self_hosted/README.md

@@ -0,0 +1,73 @@
+# Self-hosted Model API
+
+Run transcription, translation, and diarization services compatible with Reflector's GPU Model API. Works on CPU or GPU.
+
+Environment variables
+
+- REFLECTOR_GPU_APIKEY: Optional Bearer token. If unset, auth is disabled.
+- HF_TOKEN: Optional. Required for diarization to download pyannote pipelines
+
+Requirements
+
+- FFmpeg must be installed and on PATH (used for URL-based and segmented transcription)
+- Python 3.12+
+- NVIDIA GPU optional. If available, it will be used automatically
+
+Local run
+Set env vars in self_hosted/.env file
+uv sync
+
+uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+Authentication
+
+- If REFLECTOR_GPU_APIKEY is set, include header: Authorization: Bearer <key>
+
+Endpoints
+
+- POST /v1/audio/transcriptions
+
+  - multipart/form-data
+  - fields: file (single file) OR files[] (multiple files), language, batch (true/false)
+  - response: single { text, words, filename } or { results: [ ... ] }
+
+- POST /v1/audio/transcriptions-from-url
+
+  - application/json
+  - body: { audio_file_url, language, timestamp_offset }
+  - response: { text, words }
+
+- POST /translate
+
+  - text: query parameter
+  - body (application/json): { source_language, target_language }
+  - response: { text: { <src>: original, <tgt>: translated } }
+
+- POST /diarize
+  - query parameters: audio_file_url, timestamp (optional)
+  - requires HF_TOKEN to be set (for pyannote)
+  - response: { diarization: [ { start, end, speaker } ] }
+
+OpenAPI docs
+
+- Visit /docs when the server is running
+
+Docker
+
+- Not yet provided in this directory. A Dockerfile will be added later. For now, use Local run above
+
+Conformance tests
+
+# From this directory
+
+TRANSCRIPT_URL=http://localhost:8000 \
+TRANSCRIPT_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_transcript.py
+
+TRANSLATION_URL=http://localhost:8000 \
+TRANSLATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_translation.py
+
+DIARIZATION_URL=http://localhost:8000 \
+DIARIZATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_diarization.py

gpu/self_hosted/app/auth.py

@@ -0,0 +1,19 @@
+import os
+
+from fastapi import Depends, HTTPException, status
+from fastapi.security import OAuth2PasswordBearer
+
+oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+
+def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+    required_key = os.environ.get("REFLECTOR_GPU_APIKEY")
+    if not required_key:
+        return
+    if apikey == required_key:
+        return
+    raise HTTPException(
+        status_code=status.HTTP_401_UNAUTHORIZED,
+        detail="Invalid API key",
+        headers={"WWW-Authenticate": "Bearer"},
+    )

gpu/self_hosted/app/config.py

@@ -0,0 +1,12 @@
+from pathlib import Path
+
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLE_RATE = 16000
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+# App-level paths
+UPLOADS_PATH = Path("/tmp/whisper-uploads")

gpu/self_hosted/app/factory.py

@@ -0,0 +1,30 @@
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+from .routers.diarization import router as diarization_router
+from .routers.transcription import router as transcription_router
+from .routers.translation import router as translation_router
+from .services.transcriber import WhisperService
+from .services.diarizer import PyannoteDiarizationService
+from .utils import ensure_dirs
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    ensure_dirs()
+    whisper_service = WhisperService()
+    whisper_service.load()
+    app.state.whisper = whisper_service
+    diarization_service = PyannoteDiarizationService()
+    diarization_service.load()
+    app.state.diarizer = diarization_service
+    yield
+
+
+def create_app() -> FastAPI:
+    app = FastAPI(lifespan=lifespan)
+    app.include_router(transcription_router)
+    app.include_router(translation_router)
+    app.include_router(diarization_router)
+    return app

gpu/self_hosted/app/routers/diarization.py

@@ -0,0 +1,30 @@
+from typing import List
+
+from fastapi import APIRouter, Depends, Request
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.diarizer import PyannoteDiarizationService
+from ..utils import download_audio_file
+
+router = APIRouter(tags=["diarization"])
+
+
+class DiarizationSegment(BaseModel):
+    start: float
+    end: float
+    speaker: int
+
+
+class DiarizationResponse(BaseModel):
+    diarization: List[DiarizationSegment]
+
+
+@router.post(
+    "/diarize", dependencies=[Depends(apikey_auth)], response_model=DiarizationResponse
+)
+def diarize(request: Request, audio_file_url: str, timestamp: float = 0.0):
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        diarizer: PyannoteDiarizationService = request.app.state.diarizer
+        return diarizer.diarize_file(file_path, timestamp=timestamp)

gpu/self_hosted/app/routers/transcription.py

@@ -0,0 +1,109 @@
+import uuid
+from typing import Optional, Union
+
+from fastapi import APIRouter, Body, Depends, Form, HTTPException, Request, UploadFile
+from pydantic import BaseModel
+from pathlib import Path
+from ..auth import apikey_auth
+from ..config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+from ..services.transcriber import MODEL_NAME
+from ..utils import cleanup_uploaded_files, download_audio_file
+
+router = APIRouter(prefix="/v1/audio", tags=["transcription"])
+
+
+class WordTiming(BaseModel):
+    word: str
+    start: float
+    end: float
+
+
+class TranscriptResult(BaseModel):
+    text: str
+    words: list[WordTiming]
+    filename: Optional[str] = None
+
+
+class TranscriptBatchResponse(BaseModel):
+    results: list[TranscriptResult]
+
+
+@router.post(
+    "/transcriptions",
+    dependencies=[Depends(apikey_auth)],
+    response_model=Union[TranscriptResult, TranscriptBatchResponse],
+)
+def transcribe(
+    request: Request,
+    file: UploadFile = None,
+    files: list[UploadFile] | None = None,
+    model: str = Form(MODEL_NAME),
+    language: str = Form("en"),
+    batch: bool = Form(False),
+):
+    service = request.app.state.whisper
+    if not file and not files:
+        raise HTTPException(
+            status_code=400, detail="Either 'file' or 'files' parameter is required"
+        )
+    if batch and not files:
+        raise HTTPException(
+            status_code=400, detail="Batch transcription requires 'files'"
+        )
+
+    upload_files = [file] if file else files
+
+    uploaded_paths: list[Path] = []
+    with cleanup_uploaded_files(uploaded_paths):
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1].lower()
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = UPLOADS_PATH / unique_filename
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_paths.append(file_path)
+
+        if batch and len(upload_files) > 1:
+            results = []
+            for path in uploaded_paths:
+                result = service.transcribe_file(str(path), language=language)
+                result["filename"] = path.name
+                results.append(result)
+            return {"results": results}
+
+        results = []
+        for path in uploaded_paths:
+            result = service.transcribe_file(str(path), language=language)
+            result["filename"] = path.name
+            results.append(result)
+
+        return {"results": results} if len(results) > 1 else results[0]
+
+
+@router.post(
+    "/transcriptions-from-url",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranscriptResult,
+)
+def transcribe_from_url(
+    request: Request,
+    audio_file_url: str = Body(..., description="URL of the audio file to transcribe"),
+    model: str = Body(MODEL_NAME),
+    language: str = Body("en"),
+    timestamp_offset: float = Body(0.0),
+):
+    service = request.app.state.whisper
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        result = service.transcribe_vad_url_segment(
+            file_path=file_path, timestamp_offset=timestamp_offset, language=language
+        )
+        return result

gpu/self_hosted/app/routers/translation.py

@@ -0,0 +1,28 @@
+from typing import Dict
+
+from fastapi import APIRouter, Body, Depends
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.translator import TextTranslatorService
+
+router = APIRouter(tags=["translation"])
+
+translator = TextTranslatorService()
+
+
+class TranslationResponse(BaseModel):
+    text: Dict[str, str]
+
+
+@router.post(
+    "/translate",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranslationResponse,
+)
+def translate(
+    text: str,
+    source_language: str = Body("en"),
+    target_language: str = Body("fr"),
+):
+    return translator.translate(text, source_language, target_language)

gpu/self_hosted/app/services/diarizer.py

@@ -0,0 +1,42 @@
+import os
+import threading
+
+import torch
+import torchaudio
+from pyannote.audio import Pipeline
+
+
+class PyannoteDiarizationService:
+    def __init__(self):
+        self._pipeline = None
+        self._device = "cpu"
+        self._lock = threading.Lock()
+
+    def load(self):
+        self._device = "cuda" if torch.cuda.is_available() else "cpu"
+        self._pipeline = Pipeline.from_pretrained(
+            "pyannote/speaker-diarization-3.1",
+            use_auth_token=os.environ.get("HF_TOKEN"),
+        )
+        self._pipeline.to(torch.device(self._device))
+
+    def diarize_file(self, file_path: str, timestamp: float = 0.0) -> dict:
+        if self._pipeline is None:
+            self.load()
+        waveform, sample_rate = torchaudio.load(file_path)
+        with self._lock:
+            diarization = self._pipeline(
+                {"waveform": waveform, "sample_rate": sample_rate}
+            )
+        words = []
+        for diarization_segment, _, speaker in diarization.itertracks(yield_label=True):
+            words.append(
+                {
+                    "start": round(timestamp + diarization_segment.start, 3),
+                    "end": round(timestamp + diarization_segment.end, 3),
+                    "speaker": int(speaker[-2:])
+                    if speaker and speaker[-2:].isdigit()
+                    else 0,
+                }
+            )
+        return {"diarization": words}

gpu/self_hosted/app/services/transcriber.py

@@ -0,0 +1,208 @@
+import os
+import shutil
+import subprocess
+import threading
+from typing import Generator
+
+import faster_whisper
+import librosa
+import numpy as np
+import torch
+from fastapi import HTTPException
+from silero_vad import VADIterator, load_silero_vad
+
+from ..config import SAMPLE_RATE, VAD_CONFIG
+
+# Whisper configuration (service-local defaults)
+MODEL_NAME = "large-v2"
+# None delegates compute type to runtime: float16 on CUDA, int8 on CPU
+MODEL_COMPUTE_TYPE = None
+MODEL_NUM_WORKERS = 1
+CACHE_PATH = os.path.join(os.path.expanduser("~"), ".cache", "reflector-whisper")
+from ..utils import NoStdStreams
+
+
+class WhisperService:
+    def __init__(self):
+        self.model = None
+        self.device = "cpu"
+        self.lock = threading.Lock()
+
+    def load(self):
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        compute_type = MODEL_COMPUTE_TYPE or (
+            "float16" if self.device == "cuda" else "int8"
+        )
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=compute_type,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+        )
+
+    def pad_audio(self, audio_array, sample_rate: int = SAMPLE_RATE):
+        audio_duration = len(audio_array) / sample_rate
+        if audio_duration < VAD_CONFIG["silence_padding"]:
+            silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+            silence = np.zeros(silence_samples, dtype=np.float32)
+            return np.concatenate([audio_array, silence])
+        return audio_array
+
+    def enforce_word_timing_constraints(self, words: list[dict]) -> list[dict]:
+        if len(words) <= 1:
+            return words
+        enforced: list[dict] = []
+        for i, word in enumerate(words):
+            current = dict(word)
+            if i < len(words) - 1:
+                next_start = words[i + 1]["start"]
+                if current["end"] > next_start:
+                    current["end"] = next_start
+            enforced.append(current)
+        return enforced
+
+    def transcribe_file(self, file_path: str, language: str = "en") -> dict:
+        input_for_model: str | "object" = file_path
+        try:
+            audio_array, _sample_rate = librosa.load(
+                file_path, sr=SAMPLE_RATE, mono=True
+            )
+            if len(audio_array) / float(SAMPLE_RATE) < VAD_CONFIG["silence_padding"]:
+                input_for_model = self.pad_audio(audio_array, SAMPLE_RATE)
+        except Exception:
+            pass
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    input_for_model,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+        words = self.enforce_word_timing_constraints(words)
+        return {"text": text, "words": words}
+
+    def transcribe_vad_url_segment(
+        self, file_path: str, timestamp_offset: float = 0.0, language: str = "en"
+    ) -> dict:
+        def load_audio_via_ffmpeg(input_path: str, sample_rate: int) -> np.ndarray:
+            ffmpeg_bin = shutil.which("ffmpeg") or "ffmpeg"
+            cmd = [
+                ffmpeg_bin,
+                "-nostdin",
+                "-threads",
+                "1",
+                "-i",
+                input_path,
+                "-f",
+                "f32le",
+                "-acodec",
+                "pcm_f32le",
+                "-ac",
+                "1",
+                "-ar",
+                str(sample_rate),
+                "pipe:1",
+            ]
+            try:
+                proc = subprocess.run(
+                    cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True
+                )
+            except Exception as e:
+                raise HTTPException(status_code=400, detail=f"ffmpeg failed: {e}")
+            audio = np.frombuffer(proc.stdout, dtype=np.float32)
+            return audio
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLE_RATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[tuple[float, float], None, None]:
+            vad_model = load_silero_vad(onnx=False)
+            iterator = VADIterator(vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield (start / float(SAMPLE_RATE), end / float(SAMPLE_RATE))
+                    start = None
+            iterator.reset_states()
+
+        audio_array = load_audio_via_ffmpeg(file_path, SAMPLE_RATE)
+
+        merged_batches: list[tuple[float, float]] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for seg_start, seg_end in vad_segments(audio_array):
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append((batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append((batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+        for start_time, end_time in merged_batches:
+            s_idx = int(start_time * SAMPLE_RATE)
+            e_idx = int(end_time * SAMPLE_RATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = self.pad_audio(segment, SAMPLE_RATE)
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        all_words = self.enforce_word_timing_constraints(all_words)
+        return {"text": " ".join(all_text), "words": all_words}

gpu/self_hosted/app/services/translator.py

@@ -0,0 +1,44 @@
+import threading
+
+from transformers import MarianMTModel, MarianTokenizer, pipeline
+
+
+class TextTranslatorService:
+    """Simple text-to-text translator using HuggingFace MarianMT models.
+
+    This mirrors the modal translator API shape but uses text translation only.
+    """
+
+    def __init__(self):
+        self._pipeline = None
+        self._lock = threading.Lock()
+
+    def load(self, source_language: str = "en", target_language: str = "fr"):
+        # Pick a default MarianMT model pair if available; fall back to Helsinki-NLP en->fr
+        model_name = self._resolve_model_name(source_language, target_language)
+        tokenizer = MarianTokenizer.from_pretrained(model_name)
+        model = MarianMTModel.from_pretrained(model_name)
+        self._pipeline = pipeline("translation", model=model, tokenizer=tokenizer)
+
+    def _resolve_model_name(self, src: str, tgt: str) -> str:
+        # Minimal mapping; extend as needed
+        pair = (src.lower(), tgt.lower())
+        mapping = {
+            ("en", "fr"): "Helsinki-NLP/opus-mt-en-fr",
+            ("fr", "en"): "Helsinki-NLP/opus-mt-fr-en",
+            ("en", "es"): "Helsinki-NLP/opus-mt-en-es",
+            ("es", "en"): "Helsinki-NLP/opus-mt-es-en",
+            ("en", "de"): "Helsinki-NLP/opus-mt-en-de",
+            ("de", "en"): "Helsinki-NLP/opus-mt-de-en",
+        }
+        return mapping.get(pair, "Helsinki-NLP/opus-mt-en-fr")
+
+    def translate(self, text: str, source_language: str, target_language: str) -> dict:
+        if self._pipeline is None:
+            self.load(source_language, target_language)
+        with self._lock:
+            results = self._pipeline(
+                text, src_lang=source_language, tgt_lang=target_language
+            )
+        translated = results[0]["translation_text"] if results else ""
+        return {"text": {source_language: text, target_language: translated}}

gpu/self_hosted/app/utils.py

@@ -0,0 +1,107 @@
+import logging
+import os
+import sys
+import uuid
+from contextlib import contextmanager
+from typing import Mapping
+from urllib.parse import urlparse
+from pathlib import Path
+
+import requests
+from fastapi import HTTPException
+
+from .config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+
+logger = logging.getLogger(__name__)
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()
+
+
+def ensure_dirs():
+    UPLOADS_PATH.mkdir(parents=True, exist_ok=True)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> str:
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_uploads(audio_file_url: str) -> tuple[Path, str]:
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path: Path = UPLOADS_PATH / unique_filename
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    return file_path, audio_suffix
+
+
+@contextmanager
+def download_audio_file(audio_file_url: str):
+    """Download an audio file to UPLOADS_PATH and remove it after use.
+
+    Yields (file_path: Path, audio_suffix: str).
+    """
+    file_path, audio_suffix = download_audio_to_uploads(audio_file_url)
+    try:
+        yield file_path, audio_suffix
+    finally:
+        try:
+            file_path.unlink(missing_ok=True)
+        except Exception as e:
+            logger.error("Error deleting temporary file %s: %s", file_path, e)
+
+
+@contextmanager
+def cleanup_uploaded_files(file_paths: list[Path]):
+    """Ensure provided file paths are removed after use.
+
+    The provided list can be populated inside the context; all present entries
+    at exit will be deleted.
+    """
+    try:
+        yield file_paths
+    finally:
+        for path in list(file_paths):
+            try:
+                path.unlink(missing_ok=True)
+            except Exception as e:
+                logger.error("Error deleting temporary file %s: %s", path, e)

gpu/self_hosted/compose.yml

@@ -0,0 +1,10 @@
+services:
+  reflector_gpu:
+    build:
+      context: .
+    ports:
+      - "8000:8000"
+    env_file:
+      - .env
+    volumes:
+      - ./cache:/root/.cache

gpu/self_hosted/main.py

@@ -0,0 +1,3 @@
+from app.factory import create_app
+
+app = create_app()

gpu/self_hosted/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "reflector-gpu"
+version = "0.1.0"
+description = "Self-hosted GPU service for speech transcription, diarization, and translation via FastAPI."
+readme = "README.md"
+requires-python = ">=3.12"
+dependencies = [
+    "fastapi[standard]>=0.116.1",
+    "uvicorn[standard]>=0.30.0",
+    "torch>=2.3.0",
+    "faster-whisper>=1.1.0",
+    "librosa==0.10.1",
+    "numpy<2",
+    "silero-vad==5.1.0",
+    "transformers>=4.35.0",
+    "sentencepiece",
+    "pyannote.audio==3.1.0",
+    "torchaudio>=2.3.0",
+]

gpu/self_hosted/runserver.sh

@@ -0,0 +1,17 @@
+#!/bin/sh
+set -e
+
+export PATH="/root/.local/bin:$PATH"
+cd /app
+
+# Install Python dependencies at runtime (first run or when FORCE_SYNC=1)
+if [ ! -d "/app/.venv" ] || [ "$FORCE_SYNC" = "1" ]; then
+  echo "[startup] Installing Python dependencies with uv..."
+  uv sync --compile-bytecode --locked
+else
+  echo "[startup] Using existing virtual environment at /app/.venv"
+fi
+
+exec uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+

server/README.md

@@ -1,3 +1,29 @@
+## API Key Management
+
+### Finding Your User ID
+
+```bash
+# Get your OAuth sub (user ID) - requires authentication
+curl -H "Authorization: Bearer <your_jwt>" http://localhost:1250/v1/me
+# Returns: {"sub": "your-oauth-sub-here", "email": "...", ...}
+```
+
+### Creating API Keys
+
+```bash
+curl -X POST http://localhost:1250/v1/user/api-keys \
+  -H "Authorization: Bearer <your_jwt>" \
+  -H "Content-Type: application/json" \
+  -d '{"name": "My API Key"}'
+```
+
+### Using API Keys
+
+```bash
+# Use X-API-Key header instead of Authorization
+curl -H "X-API-Key: <your_api_key>" http://localhost:1250/v1/transcripts
+```
+
 ## AWS S3/SQS usage clarification
 
 Whereby.com uploads recordings directly to our S3 bucket when meetings end.

server/docs/data_retention.md

@@ -0,0 +1,95 @@
+# Data Retention and Cleanup
+
+## Overview
+
+For public instances of Reflector, a data retention policy is automatically enforced to delete anonymous user data after a configurable period (default: 7 days). This ensures compliance with privacy expectations and prevents unbounded storage growth.
+
+## Configuration
+
+### Environment Variables
+
+- `PUBLIC_MODE` (bool): Must be set to `true` to enable automatic cleanup
+- `PUBLIC_DATA_RETENTION_DAYS` (int): Number of days to retain anonymous data (default: 7)
+
+### What Gets Deleted
+
+When data reaches the retention period, the following items are automatically removed:
+
+1. **Transcripts** from anonymous users (where `user_id` is NULL):
+   - Database records
+   - Local files (audio.wav, audio.mp3, audio.json waveform)
+   - Storage files (cloud storage if configured)
+
+## Automatic Cleanup
+
+### Celery Beat Schedule
+
+When `PUBLIC_MODE=true`, a Celery beat task runs daily at 3 AM to clean up old data:
+
+```python
+# Automatically scheduled when PUBLIC_MODE=true
+"cleanup_old_public_data": {
+    "task": "reflector.worker.cleanup.cleanup_old_public_data",
+    "schedule": crontab(hour=3, minute=0),  # Daily at 3 AM
+}
+```
+
+### Running the Worker
+
+Ensure both Celery worker and beat scheduler are running:
+
+```bash
+# Start Celery worker
+uv run celery -A reflector.worker.app worker --loglevel=info
+
+# Start Celery beat scheduler (in another terminal)
+uv run celery -A reflector.worker.app beat
+```
+
+## Manual Cleanup
+
+For testing or manual intervention, use the cleanup tool:
+
+```bash
+# Delete data older than 7 days (default)
+uv run python -m reflector.tools.cleanup_old_data
+
+# Delete data older than 30 days
+uv run python -m reflector.tools.cleanup_old_data --days 30
+```
+
+Note: The manual tool uses the same implementation as the Celery worker task to ensure consistency.
+
+## Important Notes
+
+1. **User Data Deletion**: Only anonymous data (where `user_id` is NULL) is deleted. Authenticated user data is preserved.
+
+2. **Storage Cleanup**: The system properly cleans up both local files and cloud storage when configured.
+
+3. **Error Handling**: If individual deletions fail, the cleanup continues and logs errors. Failed deletions are reported in the task output.
+
+4. **Public Instance Only**: The automatic cleanup task only runs when `PUBLIC_MODE=true` to prevent accidental data loss in private deployments.
+
+## Testing
+
+Run the cleanup tests:
+
+```bash
+uv run pytest tests/test_cleanup.py -v
+```
+
+## Monitoring
+
+Check Celery logs for cleanup task execution:
+
+```bash
+# Look for cleanup task logs
+grep "cleanup_old_public_data" celery.log
+grep "Starting cleanup of old public data" celery.log
+```
+
+Task statistics are logged after each run:
+- Number of transcripts deleted
+- Number of meetings deleted
+- Number of orphaned recordings deleted
+- Any errors encountered

server/docs/gpu/api-transcription.md

@@ -0,0 +1,194 @@
+## Reflector GPU Transcription API (Specification)
+
+This document defines the Reflector GPU transcription API that all implementations must adhere to. Current implementations include NVIDIA Parakeet (NeMo) and Whisper (faster-whisper), both deployed on Modal.com. The API surface and response shapes are OpenAI/Whisper-compatible, so clients can switch implementations by changing only the base URL.
+
+### Base URL and Authentication
+
+- Example base URLs (Modal web endpoints):
+
+  - Parakeet: `https://<account>--reflector-transcriber-parakeet-web.modal.run`
+  - Whisper: `https://<account>--reflector-transcriber-web.modal.run`
+
+- All endpoints are served under `/v1` and require a Bearer token:
+
+```
+Authorization: Bearer <REFLECTOR_GPU_APIKEY>
+```
+
+Note: To switch implementations, deploy the desired variant and point `TRANSCRIPT_URL` to its base URL. The API is identical.
+
+### Supported file types
+
+`mp3, mp4, mpeg, mpga, m4a, wav, webm`
+
+### Models and languages
+
+- Parakeet (NVIDIA NeMo): default `nvidia/parakeet-tdt-0.6b-v2`
+  - Language support: only `en`. Other languages return HTTP 400.
+- Whisper (faster-whisper): default `large-v2` (or deployment-specific)
+  - Language support: multilingual (per Whisper model capabilities).
+
+Note: The `model` parameter is accepted by all implementations for interface parity. Some backends may treat it as informational.
+
+### Endpoints
+
+#### POST /v1/audio/transcriptions
+
+Transcribe one or more uploaded audio files.
+
+Request: multipart/form-data
+
+- `file` (File) — optional. Single file to transcribe.
+- `files` (File[]) — optional. One or more files to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`.
+  - Parakeet: only `en` is accepted; other values return HTTP 400
+  - Whisper: model-dependent; typically multilingual
+- `batch` (boolean) — optional, defaults to `false`.
+
+Notes:
+
+- Provide either `file` or `files`, not both. If neither is provided, HTTP 400.
+- `batch` requires `files`; using `batch=true` without `files` returns HTTP 400.
+- Response shape for multiple files is the same regardless of `batch`.
+- Files sent to this endpoint are processed in a single pass (no VAD/chunking). This is intended for short clips (roughly ≤ 30s; depends on GPU memory/model). For longer audio, prefer `/v1/audio/transcriptions-from-url` which supports VAD-based chunking.
+
+Responses
+
+Single file response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 0.0, "end": 0.5 },
+    { "word": "world", "start": 0.5, "end": 1.0 }
+  ],
+  "filename": "audio.mp3"
+}
+```
+
+Multiple files response:
+
+```json
+{
+  "results": [
+    {"filename": "a1.mp3", "text": "...", "words": [...]},
+    {"filename": "a2.mp3", "text": "...", "words": [...]}]
+}
+```
+
+Notes:
+
+- Word objects always include keys: `word`, `start`, `end`.
+- Some implementations may include a trailing space in `word` to match Whisper tokenization behavior; clients should trim if needed.
+
+Example curl (single file):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "file=@/path/to/audio.mp3" \
+  -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+Example curl (multiple files, batch):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "files=@/path/a1.mp3" -F "files=@/path/a2.mp3" \
+  -F "batch=true" -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+#### POST /v1/audio/transcriptions-from-url
+
+Transcribe a single remote audio file by URL.
+
+Request: application/json
+
+Body parameters:
+
+- `audio_file_url` (string) — required. URL of the audio file to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`. Parakeet only accepts `en`.
+- `timestamp_offset` (number) — optional, defaults to `0.0`. Added to each word's `start`/`end` in the response.
+
+```json
+{
+  "audio_file_url": "https://example.com/audio.mp3",
+  "model": "nvidia/parakeet-tdt-0.6b-v2",
+  "language": "en",
+  "timestamp_offset": 0.0
+}
+```
+
+Response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 10.0, "end": 10.5 },
+    { "word": "world", "start": 10.5, "end": 11.0 }
+  ]
+}
+```
+
+Notes:
+
+- `timestamp_offset` is added to each word’s `start`/`end` in the response.
+- Implementations may perform VAD-based chunking and batching for long-form audio; word timings are adjusted accordingly.
+
+Example curl:
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+        "audio_file_url": "https://example.com/audio.mp3",
+        "language": "en",
+        "timestamp_offset": 0
+      }' \
+  "$BASE_URL/v1/audio/transcriptions-from-url"
+```
+
+### Error handling
+
+- 400 Bad Request
+  - Parakeet: `language` other than `en`
+  - Missing required parameters (`file`/`files` for upload; `audio_file_url` for URL endpoint)
+  - Unsupported file extension
+- 401 Unauthorized
+  - Missing or invalid Bearer token
+- 404 Not Found
+  - `audio_file_url` does not exist
+
+### Implementation details
+
+- GPUs: A10G for small-file/live, L40S for large-file URL transcription (subject to deployment)
+- VAD chunking and segment batching; word timings adjusted and overlapping ends constrained
+- Pads very short segments (< 0.5s) to avoid model crashes on some backends
+
+### Server configuration (Reflector API)
+
+Set the Reflector server to use the Modal backend and point `TRANSCRIPT_URL` to your chosen deployment:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://<account>--reflector-transcriber-parakeet-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=<REFLECTOR_GPU_APIKEY>
+```
+
+### Conformance tests
+
+Use the pytest-based conformance tests to validate any new implementation (including self-hosted) against this spec:
+
+```
+TRANSCRIPT_URL=https://<your-deployment-base> \
+TRANSCRIPT_MODAL_API_KEY=your-api-key \
+uv run -m pytest -m model_api --no-cov server/tests/test_model_api_transcript.py
+```

server/docs/video-platforms/README.md

@@ -0,0 +1,236 @@
+# Reflector Architecture: Whereby + Daily.co Recording Storage
+
+## System Overview
+
+```mermaid
+graph TB
+    subgraph "Actors"
+        APP[Our App<br/>Reflector]
+        WHEREBY[Whereby Service<br/>External]
+        DAILY[Daily.co Service<br/>External]
+    end
+
+    subgraph "AWS S3 Buckets"
+        TRANSCRIPT_BUCKET[Transcript Bucket<br/>reflector-transcripts<br/>Output: Processed MP3s]
+        WHEREBY_BUCKET[Whereby Bucket<br/>reflector-whereby-recordings<br/>Input: Raw MP4s]
+        DAILY_BUCKET[Daily.co Bucket<br/>reflector-dailyco-recordings<br/>Input: Raw WebM tracks]
+    end
+
+    subgraph "AWS Infrastructure"
+        SQS[SQS Queue<br/>Whereby notifications]
+    end
+
+    subgraph "Database"
+        DB[(PostgreSQL<br/>Recordings, Transcripts, Meetings)]
+    end
+
+    APP -->|Write processed| TRANSCRIPT_BUCKET
+    APP -->|Read/Delete| WHEREBY_BUCKET
+    APP -->|Read/Delete| DAILY_BUCKET
+    APP -->|Poll| SQS
+    APP -->|Store metadata| DB
+
+    WHEREBY -->|Write recordings| WHEREBY_BUCKET
+    WHEREBY_BUCKET -->|S3 Event| SQS
+    WHEREBY -->|Participant webhooks<br/>room.client.joined/left| APP
+
+    DAILY -->|Write recordings| DAILY_BUCKET
+    DAILY -->|Recording webhook<br/>recording.ready-to-download| APP
+```
+
+**Note on Webhook vs S3 Event for Recording Processing:**
+- **Whereby**: Uses S3 Events → SQS for recording availability (S3 as source of truth, no race conditions)
+- **Daily.co**: Uses webhooks for recording availability (more immediate, built-in reliability)
+- **Both**: Use webhooks for participant tracking (real-time updates)
+
+## Credentials & Permissions
+
+```mermaid
+graph LR
+    subgraph "Master Credentials"
+        MASTER[TRANSCRIPT_STORAGE_AWS_*<br/>Access Key ID + Secret]
+    end
+
+    subgraph "Whereby Upload Credentials"
+        WHEREBY_CREDS[AWS_WHEREBY_ACCESS_KEY_*<br/>Access Key ID + Secret]
+    end
+
+    subgraph "Daily.co Upload Role"
+        DAILY_ROLE[DAILY_STORAGE_AWS_ROLE_ARN<br/>IAM Role ARN]
+    end
+
+    subgraph "Our App Uses"
+        MASTER -->|Read/Write/Delete| TRANSCRIPT_BUCKET[Transcript Bucket]
+        MASTER -->|Read/Delete| WHEREBY_BUCKET[Whereby Bucket]
+        MASTER -->|Read/Delete| DAILY_BUCKET[Daily.co Bucket]
+        MASTER -->|Poll/Delete| SQS[SQS Queue]
+    end
+
+    subgraph "We Give To Services"
+        WHEREBY_CREDS -->|Passed in API call| WHEREBY_SERVICE[Whereby Service]
+        WHEREBY_SERVICE -->|Write Only| WHEREBY_BUCKET
+
+        DAILY_ROLE -->|Passed in API call| DAILY_SERVICE[Daily.co Service]
+        DAILY_SERVICE -->|Assume Role| DAILY_ROLE
+        DAILY_SERVICE -->|Write Only| DAILY_BUCKET
+    end
+```
+
+# Video Platform Recording Integration
+
+This document explains how Reflector receives and identifies multitrack audio recordings from different video platforms.
+
+## Platform Comparison
+
+| Platform | Delivery Method | Track Identification |
+|----------|----------------|---------------------|
+| **Daily.co** | Webhook | Explicit track list in payload |
+| **Whereby** | SQS (S3 notifications) | Single file per notification |
+
+---
+
+## Daily.co
+
+**Note:** Primary discovery via polling (`poll_daily_recordings`), webhooks as backup.
+
+Daily.co uses **webhooks** to notify Reflector when recordings are ready.
+
+### How It Works
+
+1. **Daily.co sends webhook** when recording is ready
+   - Event type: `recording.ready-to-download`
+   - Endpoint: `/v1/daily/webhook` (`reflector/views/daily.py:46-102`)
+
+2. **Webhook payload explicitly includes track list**:
+```json
+{
+  "recording_id": "7443ee0a-dab1-40eb-b316-33d6c0d5ff88",
+  "room_name": "daily-20251020193458",
+  "tracks": [
+    {
+      "type": "audio",
+      "s3Key": "monadical/daily-20251020193458/1760988935484-52f7f48b-fbab-431f-9a50-87b9abfc8255-cam-audio-1760988935922",
+      "size": 831843
+    },
+    {
+      "type": "audio",
+      "s3Key": "monadical/daily-20251020193458/1760988935484-a37c35e3-6f8e-4274-a482-e9d0f102a732-cam-audio-1760988943823",
+      "size": 408438
+    },
+    {
+      "type": "video",
+      "s3Key": "monadical/daily-20251020193458/...-video.webm",
+      "size": 30000000
+    }
+  ]
+}
+```
+
+3. **System extracts audio tracks** (`daily.py:211`):
+```python
+track_keys = [t.s3Key for t in tracks if t.type == "audio"]
+```
+
+4. **Triggers multitrack processing** (`daily.py:213-218`):
+```python
+process_multitrack_recording.delay(
+    bucket_name=bucket_name,  # reflector-dailyco-local
+    room_name=room_name,      # daily-20251020193458
+    recording_id=recording_id, # 7443ee0a-dab1-40eb-b316-33d6c0d5ff88
+    track_keys=track_keys      # Only audio s3Keys
+)
+```
+
+### Key Advantage: No Ambiguity
+
+Even though multiple meetings may share the same S3 bucket/folder (`monadical/`), **there's no ambiguity** because:
+- Each webhook payload contains the exact `s3Key` list for that specific `recording_id`
+- No need to scan folders or guess which files belong together
+- Each track's s3Key includes the room timestamp subfolder (e.g., `daily-20251020193458/`)
+
+The room name includes timestamp (`daily-20251020193458`) to keep recordings organized, but **the webhook's explicit track list is what prevents mixing files from different meetings**.
+
+### Track Timeline Extraction
+
+Daily.co provides timing information in two places:
+
+**1. PyAV WebM Metadata (current approach)**:
+```python
+# Read from WebM container stream metadata
+stream.start_time = 8.130s  # Meeting-relative timing
+```
+
+**2. Filename Timestamps (alternative approach, commit 3bae9076)**:
+```
+Filename format: {recording_start_ts}-{uuid}-cam-audio-{track_start_ts}.webm
+Example: 1760988935484-52f7f48b-fbab-431f-9a50-87b9abfc8255-cam-audio-1760988935922.webm
+
+Parse timestamps:
+- recording_start_ts: 1760988935484 (Unix ms)
+- track_start_ts: 1760988935922 (Unix ms)
+- offset: (1760988935922 - 1760988935484) / 1000 = 0.438s
+```
+
+**Time Difference (PyAV vs Filename)**:
+```
+Track 0:
+  Filename offset: 438ms
+  PyAV metadata:   229ms
+  Difference:      209ms
+
+Track 1:
+  Filename offset: 8339ms
+  PyAV metadata:   8130ms
+  Difference:      209ms
+```
+
+**Consistent 209ms delta** suggests network/encoding delay between file upload initiation (filename) and actual audio stream start (metadata).
+
+**Current implementation uses PyAV metadata** because:
+- More accurate (represents when audio actually started)
+- Padding BEFORE transcription produces correct Whisper timestamps automatically
+- No manual offset adjustment needed during transcript merge
+
+### Why Re-encoding During Padding
+
+Padding coincidentally involves re-encoding, which is important for Daily.co + Whisper:
+
+**Problem:** Daily.co skips frames in recordings when microphone is muted or paused
+- WebM containers have gaps where audio frames should be
+- Whisper doesn't understand these gaps and produces incorrect timestamps
+- Example: 5s of audio with 2s muted → file has frames only for 3s, Whisper thinks duration is 3s
+
+**Solution:** Re-encoding via PyAV filter graph (`adelay` + `aresample`)
+- Restores missing frames as silence
+- Produces continuous audio stream without gaps
+- Whisper now sees correct duration and produces accurate timestamps
+
+**Why combined with padding:**
+- Already re-encoding for padding (adding initial silence)
+- More performant to do both operations in single PyAV pipeline
+- Padded values needed for mixdown anyway (creating final MP3)
+
+Implementation: `main_multitrack_pipeline.py:_apply_audio_padding_streaming()`
+
+---
+
+## Whereby (SQS-based)
+
+Whereby uses **AWS SQS** (via S3 notifications) to notify Reflector when files are uploaded.
+
+### How It Works
+
+1. **Whereby uploads recording** to S3
+2. **S3 sends notification** to SQS queue (one notification per file)
+3. **Reflector polls SQS queue** (`worker/process.py:process_messages()`)
+4. **System processes single file** (`worker/process.py:process_recording()`)
+
+### Key Difference from Daily.co
+
+**Whereby (SQS):** System receives S3 notification "file X was created" - only knows about one file at a time, would need to scan folder to find related files
+
+**Daily.co (Webhook):** Daily explicitly tells system which files belong together in the webhook payload
+
+---
+
+

server/docs/webhook.md

@@ -0,0 +1,233 @@
+# Reflector Webhook Documentation
+
+## Overview
+
+Reflector supports webhook notifications to notify external systems when transcript processing is completed. Webhooks can be configured per room and are triggered automatically after a transcript is successfully processed.
+
+## Configuration
+
+Webhooks are configured at the room level with two fields:
+- `webhook_url`: The HTTPS endpoint to receive webhook notifications
+- `webhook_secret`: Optional secret key for HMAC signature verification (auto-generated if not provided)
+
+## Events
+
+### `transcript.completed`
+
+Triggered when a transcript has been fully processed, including transcription, diarization, summarization, topic detection and calendar event integration.
+
+### `test`
+
+A test event that can be triggered manually to verify webhook configuration.
+
+## Webhook Request Format
+
+### Headers
+
+All webhook requests include the following headers:
+
+| Header | Description | Example |
+|--------|-------------|---------|
+| `Content-Type` | Always `application/json` | `application/json` |
+| `User-Agent` | Identifies Reflector as the source | `Reflector-Webhook/1.0` |
+| `X-Webhook-Event` | The event type | `transcript.completed` or `test` |
+| `X-Webhook-Retry` | Current retry attempt number | `0`, `1`, `2`... |
+| `X-Webhook-Signature` | HMAC signature (if secret configured) | `t=1735306800,v1=abc123...` |
+
+### Signature Verification
+
+If a webhook secret is configured, Reflector includes an HMAC-SHA256 signature in the `X-Webhook-Signature` header to verify the webhook authenticity.
+
+The signature format is: `t={timestamp},v1={signature}`
+
+To verify the signature:
+1. Extract the timestamp and signature from the header
+2. Create the signed payload: `{timestamp}.{request_body}`
+3. Compute HMAC-SHA256 of the signed payload using your webhook secret
+4. Compare the computed signature with the received signature
+
+Example verification (Python):
+```python
+import hmac
+import hashlib
+
+def verify_webhook_signature(payload: bytes, signature_header: str, secret: str) -> bool:
+    # Parse header: "t=1735306800,v1=abc123..."
+    parts = dict(part.split("=") for part in signature_header.split(","))
+    timestamp = parts["t"]
+    received_signature = parts["v1"]
+
+    # Create signed payload
+    signed_payload = f"{timestamp}.{payload.decode('utf-8')}"
+
+    # Compute expected signature
+    expected_signature = hmac.new(
+        secret.encode("utf-8"),
+        signed_payload.encode("utf-8"),
+        hashlib.sha256
+    ).hexdigest()
+
+    # Compare signatures
+    return hmac.compare_digest(expected_signature, received_signature)
+```
+
+## Event Payloads
+
+### `transcript.completed` Event
+
+This event includes a convenient URL for accessing the transcript:
+- `frontend_url`: Direct link to view the transcript in the web interface
+
+```json
+{
+  "event": "transcript.completed",
+  "event_id": "transcript.completed-abc-123-def-456",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "transcript": {
+    "id": "abc-123-def-456",
+    "room_id": "room-789",
+    "created_at": "2025-08-27T12:00:00Z",
+    "duration": 1800.5,
+    "title": "Q3 Product Planning Meeting",
+    "short_summary": "Team discussed Q3 product roadmap, prioritizing mobile app features and API improvements.",
+    "long_summary": "The product team met to finalize the Q3 roadmap. Key decisions included...",
+    "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone to today's meeting...",
+    "topics": [
+      {
+        "title": "Introduction and Agenda",
+        "summary": "Meeting kickoff with agenda review",
+        "timestamp": 0.0,
+        "duration": 120.0,
+        "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone..."
+      },
+      {
+        "title": "Mobile App Features Discussion",
+        "summary": "Team reviewed proposed mobile app features for Q3",
+        "timestamp": 120.0,
+        "duration": 600.0,
+        "webvtt": "WEBVTT\n\n00:02:00.000 --> 00:02:10.000\n<v Speaker 2>Let's talk about the mobile app..."
+      }
+    ],
+    "participants": [
+      {
+        "id": "participant-1",
+        "name": "John Doe",
+        "speaker": "Speaker 1"
+      },
+      {
+        "id": "participant-2",
+        "name": "Jane Smith",
+        "speaker": "Speaker 2"
+      }
+    ],
+    "source_language": "en",
+    "target_language": "en",
+    "status": "completed",
+    "frontend_url": "https://app.reflector.com/transcripts/abc-123-def-456"
+  },
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  },
+  "calendar_event": {
+    "id": "calendar-event-123",
+    "ics_uid": "event-123",
+    "title": "Q3 Product Planning Meeting",
+    "start_time": "2025-08-27T12:00:00Z",
+    "end_time": "2025-08-27T12:30:00Z",
+    "description": "Team discussed Q3 product roadmap, prioritizing mobile app features and API improvements.",
+    "location": "Conference Room 1",
+    "attendees": [
+      {
+        "id": "participant-1",
+        "name": "John Doe",
+        "speaker": "Speaker 1"
+      },
+      {
+        "id": "participant-2",
+        "name": "Jane Smith",
+        "speaker": "Speaker 2"
+      }
+    ]
+  }
+}
+```
+
+### `test` Event
+
+```json
+{
+  "event": "test",
+  "event_id": "test.2025-08-27T12:34:56.789012Z",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "message": "This is a test webhook from Reflector",
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  }
+}
+```
+
+## Retry Policy
+
+Webhooks are delivered with automatic retry logic to handle transient failures. When a webhook delivery fails due to server errors or network issues, Reflector will automatically retry the delivery multiple times over an extended period.
+
+### Retry Mechanism
+
+Reflector implements an exponential backoff strategy for webhook retries:
+
+- **Initial retry delay**: 60 seconds after the first failure
+- **Exponential backoff**: Each subsequent retry waits approximately twice as long as the previous one
+- **Maximum retry interval**: 1 hour (backoff is capped at this duration)
+- **Maximum retry attempts**: 30 attempts total
+- **Total retry duration**: Retries continue for approximately 24 hours
+
+### How Retries Work
+
+When a webhook fails, Reflector will:
+1. Wait 60 seconds, then retry (attempt #1)
+2. If it fails again, wait ~2 minutes, then retry (attempt #2)
+3. Continue doubling the wait time up to a maximum of 1 hour between attempts
+4. Keep retrying at 1-hour intervals until successful or 30 attempts are exhausted
+
+The `X-Webhook-Retry` header indicates the current retry attempt number (0 for the initial attempt, 1 for first retry, etc.), allowing your endpoint to track retry attempts.
+
+### Retry Behavior by HTTP Status Code
+
+| Status Code | Behavior |
+|-------------|----------|
+| 2xx (Success) | No retry, webhook marked as delivered |
+| 4xx (Client Error) | No retry, request is considered permanently failed |
+| 5xx (Server Error) | Automatic retry with exponential backoff |
+| Network/Timeout Error | Automatic retry with exponential backoff |
+
+**Important Notes:**
+- Webhooks timeout after 30 seconds. If your endpoint takes longer to respond, it will be considered a timeout error and retried.
+- During the retry period (~24 hours), you may receive the same webhook multiple times if your endpoint experiences intermittent failures.
+- There is no mechanism to manually retry failed webhooks after the retry period expires.
+
+## Testing Webhooks
+
+You can test your webhook configuration before processing transcripts:
+
+```http
+POST /v1/rooms/{room_id}/webhook/test
+```
+
+Response:
+```json
+{
+  "success": true,
+  "status_code": 200,
+  "message": "Webhook test successful",
+  "response_preview": "OK"
+}
+```
+
+Or in case of failure:
+```json
+{
+  "success": false,
+  "error": "Webhook request timed out (10 seconds)"
+}
+```

server/env.example

@@ -27,7 +27,7 @@ AUTH_JWT_AUDIENCE=
 #TRANSCRIPT_MODAL_API_KEY=xxxxx
 
 TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
+TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-parakeet-web.modal.run
 TRANSCRIPT_MODAL_API_KEY=
 
 ## =======================================================
@@ -71,3 +71,30 @@ DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
 
 ## Sentry DSN configuration
 #SENTRY_DSN=
+
+## =======================================================
+## Video Platform Configuration
+## =======================================================
+
+## Whereby
+#WHEREBY_API_KEY=your-whereby-api-key
+#WHEREBY_WEBHOOK_SECRET=your-whereby-webhook-secret
+#WHEREBY_STORAGE_AWS_ACCESS_KEY_ID=your-aws-key
+#WHEREBY_STORAGE_AWS_SECRET_ACCESS_KEY=your-aws-secret
+#AWS_PROCESS_RECORDING_QUEUE_URL=https://sqs.us-west-2.amazonaws.com/...
+
+## Daily.co
+#DAILY_API_KEY=your-daily-api-key
+#DAILY_WEBHOOK_SECRET=your-daily-webhook-secret
+#DAILY_SUBDOMAIN=your-subdomain
+#DAILY_WEBHOOK_UUID=  # Auto-populated by recreate_daily_webhook.py script
+#DAILYCO_STORAGE_AWS_ROLE_ARN=...  # IAM role ARN for Daily.co S3 access
+#DAILYCO_STORAGE_AWS_BUCKET_NAME=reflector-dailyco
+#DAILYCO_STORAGE_AWS_REGION=us-west-2
+
+## Whereby (optional separate bucket)
+#WHEREBY_STORAGE_AWS_BUCKET_NAME=reflector-whereby
+#WHEREBY_STORAGE_AWS_REGION=us-east-1
+
+## Platform Configuration
+#DEFAULT_VIDEO_PLATFORM=whereby          # Default platform for new rooms

server/gpu/modal_deployments/reflector_transcriber.py

@@ -1,161 +0,0 @@
-import os
-import tempfile
-import threading
-
-import modal
-from pydantic import BaseModel
-
-MODELS_DIR = "/models"
-
-MODEL_NAME = "large-v2"
-MODEL_COMPUTE_TYPE: str = "float16"
-MODEL_NUM_WORKERS: int = 1
-
-MINUTES = 60  # seconds
-
-volume = modal.Volume.from_name("models", create_if_missing=True)
-
-app = modal.App("reflector-transcriber")
-
-
-def download_model():
-    from faster_whisper import download_model
-
-    volume.reload()
-
-    download_model(MODEL_NAME, cache_dir=MODELS_DIR)
-
-    volume.commit()
-
-
-image = (
-    modal.Image.debian_slim(python_version="3.12")
-    .pip_install(
-        "huggingface_hub==0.27.1",
-        "hf-transfer==0.1.9",
-        "torch==2.5.1",
-        "faster-whisper==1.1.1",
-    )
-    .env(
-        {
-            "HF_HUB_ENABLE_HF_TRANSFER": "1",
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
-            ),
-        }
-    )
-    .run_function(download_model, volumes={MODELS_DIR: volume})
-)
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=5 * MINUTES,
-    scaledown_window=5 * MINUTES,
-    allow_concurrent_inputs=6,
-    image=image,
-    volumes={MODELS_DIR: volume},
-)
-class Transcriber:
-    @modal.enter()
-    def enter(self):
-        import faster_whisper
-        import torch
-
-        self.lock = threading.Lock()
-        self.use_gpu = torch.cuda.is_available()
-        self.device = "cuda" if self.use_gpu else "cpu"
-        self.model = faster_whisper.WhisperModel(
-            MODEL_NAME,
-            device=self.device,
-            compute_type=MODEL_COMPUTE_TYPE,
-            num_workers=MODEL_NUM_WORKERS,
-            download_root=MODELS_DIR,
-            local_files_only=True,
-        )
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        audio_data: str,
-        audio_suffix: str,
-        language: str,
-    ):
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
-
-            with self.lock:
-                segments, _ = self.model.transcribe(
-                    fp.name,
-                    language=language,
-                    beam_size=5,
-                    word_timestamps=True,
-                    vad_filter=True,
-                    vad_parameters={"min_silence_duration_ms": 500},
-                )
-
-            segments = list(segments)
-            text = "".join(segment.text for segment in segments)
-            words = [
-                {"word": word.word, "start": word.start, "end": word.end}
-                for segment in segments
-                for word in segment.words
-            ]
-
-            return {"text": text, "words": words}
-
-
-@app.function(
-    scaledown_window=60,
-    timeout=60,
-    allow_concurrent_inputs=40,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-    volumes={MODELS_DIR: volume},
-)
-@modal.asgi_app()
-def web():
-    from fastapi import Body, Depends, FastAPI, HTTPException, UploadFile, status
-    from fastapi.security import OAuth2PasswordBearer
-    from typing_extensions import Annotated
-
-    transcriber = Transcriber()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    supported_file_types = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    class TranscriptResponse(BaseModel):
-        result: dict
-
-    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
-    def transcribe(
-        file: UploadFile,
-        model: str = "whisper-1",
-        language: Annotated[str, Body(...)] = "en",
-    ) -> TranscriptResponse:
-        audio_data = file.file.read()
-        audio_suffix = file.filename.split(".")[-1]
-        assert audio_suffix in supported_file_types
-
-        func = transcriber.transcribe_segment.spawn(
-            audio_data=audio_data,
-            audio_suffix=audio_suffix,
-            language=language,
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_transcriber_parakeet_v3.py

@@ -1,622 +0,0 @@
-import logging
-import os
-import sys
-import threading
-import uuid
-from typing import Mapping, NewType
-from urllib.parse import urlparse
-
-import modal
-
-MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v3"
-SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
-SAMPLERATE = 16000
-UPLOADS_PATH = "/uploads"
-CACHE_PATH = "/cache"
-VAD_CONFIG = {
-    "max_segment_duration": 30.0,
-    "batch_max_files": 10,
-    "batch_max_duration": 5.0,
-    "min_segment_duration": 0.02,
-    "silence_padding": 0.5,
-    "window_size": 512,
-}
-
-ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
-AudioFileExtension = NewType("AudioFileExtension", str)
-
-app = modal.App("reflector-transcriber-parakeet-v3")
-
-# Volume for caching model weights
-model_cache = modal.Volume.from_name("parakeet-model-cache", create_if_missing=True)
-# Volume for temporary file uploads
-upload_volume = modal.Volume.from_name("parakeet-uploads", create_if_missing=True)
-
-image = (
-    modal.Image.from_registry(
-        "nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04", add_python="3.12"
-    )
-    .env(
-        {
-            "HF_HUB_ENABLE_HF_TRANSFER": "1",
-            "HF_HOME": "/cache",
-            "DEBIAN_FRONTEND": "noninteractive",
-            "CXX": "g++",
-            "CC": "g++",
-        }
-    )
-    .apt_install("ffmpeg")
-    .pip_install(
-        "hf_transfer==0.1.9",
-        "huggingface_hub[hf-xet]==0.31.2",
-        "nemo_toolkit[asr]==2.3.0",
-        "cuda-python==12.8.0",
-        "fastapi==0.115.12",
-        "numpy<2",
-        "librosa==0.10.1",
-        "requests",
-        "silero-vad==5.1.0",
-        "torch",
-    )
-    .entrypoint([])  # silence chatty logs by container on start
-)
-
-
-def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
-    parsed_url = urlparse(url)
-    url_path = parsed_url.path
-
-    for ext in SUPPORTED_FILE_EXTENSIONS:
-        if url_path.lower().endswith(f".{ext}"):
-            return AudioFileExtension(ext)
-
-    content_type = headers.get("content-type", "").lower()
-    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
-        return AudioFileExtension("mp3")
-    if "audio/wav" in content_type:
-        return AudioFileExtension("wav")
-    if "audio/mp4" in content_type:
-        return AudioFileExtension("mp4")
-
-    raise ValueError(
-        f"Unsupported audio format for URL: {url}. "
-        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
-    )
-
-
-def download_audio_to_volume(
-    audio_file_url: str,
-) -> tuple[ParakeetUniqFilename, AudioFileExtension]:
-    import requests
-    from fastapi import HTTPException
-
-    response = requests.head(audio_file_url, allow_redirects=True)
-    if response.status_code == 404:
-        raise HTTPException(status_code=404, detail="Audio file not found")
-
-    response = requests.get(audio_file_url, allow_redirects=True)
-    response.raise_for_status()
-
-    audio_suffix = detect_audio_format(audio_file_url, response.headers)
-    unique_filename = ParakeetUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
-    file_path = f"{UPLOADS_PATH}/{unique_filename}"
-
-    with open(file_path, "wb") as f:
-        f.write(response.content)
-
-    upload_volume.commit()
-    return unique_filename, audio_suffix
-
-
-def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
-    """Add 0.5 seconds of silence if audio is less than 500ms.
-
-    This is a workaround for a Parakeet bug where very short audio (<500ms) causes:
-    ValueError: `char_offsets`: [] and `processed_tokens`: [157, 834, 834, 841]
-    have to be of the same length
-
-    See: https://github.com/NVIDIA/NeMo/issues/8451
-    """
-    import numpy as np
-
-    audio_duration = len(audio_array) / sample_rate
-    if audio_duration < 0.5:
-        silence_samples = int(sample_rate * 0.5)
-        silence = np.zeros(silence_samples, dtype=np.float32)
-        return np.concatenate([audio_array, silence])
-    return audio_array
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=600,
-    scaledown_window=300,
-    image=image,
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    enable_memory_snapshot=True,
-    experimental_options={"enable_gpu_snapshot": True},
-)
-@modal.concurrent(max_inputs=10)
-class TranscriberParakeetLive:
-    @modal.enter(snap=True)
-    def enter(self):
-        import nemo.collections.asr as nemo_asr
-
-        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
-
-        self.lock = threading.Lock()
-        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
-        device = next(self.model.parameters()).device
-        print(f"Model is on device: {device}")
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        filename: str,
-    ):
-        import librosa
-
-        upload_volume.reload()
-
-        file_path = f"{UPLOADS_PATH}/{filename}"
-        if not os.path.exists(file_path):
-            raise FileNotFoundError(f"File not found: {file_path}")
-
-        audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-        padded_audio = pad_audio(audio_array, sample_rate)
-
-        with self.lock:
-            with NoStdStreams():
-                (output,) = self.model.transcribe([padded_audio], timestamps=True)
-
-        text = output.text.strip()
-        words = [
-            {
-                "word": word_info["word"] + " ",
-                "start": round(word_info["start"], 2),
-                "end": round(word_info["end"], 2),
-            }
-            for word_info in output.timestamp["word"]
-        ]
-
-        return {"text": text, "words": words}
-
-    @modal.method()
-    def transcribe_batch(
-        self,
-        filenames: list[str],
-    ):
-        import librosa
-
-        upload_volume.reload()
-
-        results = []
-        audio_arrays = []
-
-        # Load all audio files with padding
-        for filename in filenames:
-            file_path = f"{UPLOADS_PATH}/{filename}"
-            if not os.path.exists(file_path):
-                raise FileNotFoundError(f"Batch file not found: {file_path}")
-
-            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-            padded_audio = pad_audio(audio_array, sample_rate)
-            audio_arrays.append(padded_audio)
-
-        with self.lock:
-            with NoStdStreams():
-                outputs = self.model.transcribe(audio_arrays, timestamps=True)
-
-        # Process results for each file
-        for i, (filename, output) in enumerate(zip(filenames, outputs)):
-            text = output.text.strip()
-
-            words = [
-                {
-                    "word": word_info["word"] + " ",
-                    "start": round(word_info["start"], 2),
-                    "end": round(word_info["end"], 2),
-                }
-                for word_info in output.timestamp["word"]
-            ]
-
-            results.append(
-                {
-                    "filename": filename,
-                    "text": text,
-                    "words": words,
-                }
-            )
-
-        return results
-
-
-# L40S class for file transcription (bigger files)
-@app.cls(
-    gpu="L40S",
-    timeout=900,
-    image=image,
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    enable_memory_snapshot=True,
-    experimental_options={"enable_gpu_snapshot": True},
-)
-class TranscriberParakeetFile:
-    @modal.enter(snap=True)
-    def enter(self):
-        import nemo.collections.asr as nemo_asr
-        import torch
-        from silero_vad import load_silero_vad
-
-        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
-
-        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
-        device = next(self.model.parameters()).device
-        print(f"Model is on device: {device}")
-
-        torch.set_num_threads(1)
-        self.vad_model = load_silero_vad(onnx=False)
-        print("Silero VAD initialized")
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        filename: str,
-        timestamp_offset: float = 0.0,
-    ):
-        import librosa
-        import numpy as np
-        from silero_vad import VADIterator
-
-        def load_and_convert_audio(file_path):
-            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
-            return audio_array
-
-        def vad_segment_generator(audio_array):
-            """Generate speech segments using VAD with start/end sample indices"""
-            vad_iterator = VADIterator(self.vad_model, sampling_rate=SAMPLERATE)
-            window_size = VAD_CONFIG["window_size"]
-            start = None
-
-            for i in range(0, len(audio_array), window_size):
-                chunk = audio_array[i : i + window_size]
-                if len(chunk) < window_size:
-                    chunk = np.pad(
-                        chunk, (0, window_size - len(chunk)), mode="constant"
-                    )
-
-                speech_dict = vad_iterator(chunk)
-                if not speech_dict:
-                    continue
-
-                if "start" in speech_dict:
-                    start = speech_dict["start"]
-                    continue
-
-                if "end" in speech_dict and start is not None:
-                    end = speech_dict["end"]
-                    start_time = start / float(SAMPLERATE)
-                    end_time = end / float(SAMPLERATE)
-
-                    # Extract the actual audio segment
-                    audio_segment = audio_array[start:end]
-
-                    yield (start_time, end_time, audio_segment)
-                    start = None
-
-            vad_iterator.reset_states()
-
-        def vad_segment_filter(segments):
-            """Filter VAD segments by duration and chunk large segments"""
-            min_dur = VAD_CONFIG["min_segment_duration"]
-            max_dur = VAD_CONFIG["max_segment_duration"]
-
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
-
-                # Skip very small segments
-                if segment_duration < min_dur:
-                    continue
-
-                # If segment is within max duration, yield as-is
-                if segment_duration <= max_dur:
-                    yield (start_time, end_time, audio_segment)
-                    continue
-
-                # Chunk large segments into smaller pieces
-                chunk_samples = int(max_dur * SAMPLERATE)
-                current_start = start_time
-
-                for chunk_offset in range(0, len(audio_segment), chunk_samples):
-                    chunk_audio = audio_segment[
-                        chunk_offset : chunk_offset + chunk_samples
-                    ]
-                    if len(chunk_audio) == 0:
-                        break
-
-                    chunk_duration = len(chunk_audio) / float(SAMPLERATE)
-                    chunk_end = current_start + chunk_duration
-
-                    # Only yield chunks that meet minimum duration
-                    if chunk_duration >= min_dur:
-                        yield (current_start, chunk_end, chunk_audio)
-
-                    current_start = chunk_end
-
-        def batch_segments(segments, max_files=10, max_duration=5.0):
-            batch = []
-            batch_duration = 0.0
-
-            for start_time, end_time, audio_segment in segments:
-                segment_duration = end_time - start_time
-
-                if segment_duration < VAD_CONFIG["silence_padding"]:
-                    silence_samples = int(
-                        (VAD_CONFIG["silence_padding"] - segment_duration) * SAMPLERATE
-                    )
-                    padding = np.zeros(silence_samples, dtype=np.float32)
-                    audio_segment = np.concatenate([audio_segment, padding])
-                    segment_duration = VAD_CONFIG["silence_padding"]
-
-                batch.append((start_time, end_time, audio_segment))
-                batch_duration += segment_duration
-
-                if len(batch) >= max_files or batch_duration >= max_duration:
-                    yield batch
-                    batch = []
-                    batch_duration = 0.0
-
-            if batch:
-                yield batch
-
-        def transcribe_batch(model, audio_segments):
-            with NoStdStreams():
-                outputs = model.transcribe(audio_segments, timestamps=True)
-            return outputs
-
-        def emit_results(
-            results,
-            segments_info,
-            batch_index,
-            total_batches,
-        ):
-            """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
-            for i, (output, (start_time, end_time, _)) in enumerate(
-                zip(results, segments_info)
-            ):
-                text = output.text.strip()
-                words = [
-                    {
-                        "word": word_info["word"],
-                        "start": round(
-                            word_info["start"] + start_time + timestamp_offset, 2
-                        ),
-                        "end": round(
-                            word_info["end"] + start_time + timestamp_offset, 2
-                        ),
-                    }
-                    for word_info in output.timestamp["word"]
-                ]
-
-                yield text, words
-
-        upload_volume.reload()
-
-        file_path = f"{UPLOADS_PATH}/{filename}"
-        if not os.path.exists(file_path):
-            raise FileNotFoundError(f"File not found: {file_path}")
-
-        audio_array = load_and_convert_audio(file_path)
-        total_duration = len(audio_array) / float(SAMPLERATE)
-        processed_duration = 0.0
-
-        all_text_parts = []
-        all_words = []
-
-        raw_segments = vad_segment_generator(audio_array)
-        filtered_segments = vad_segment_filter(raw_segments)
-        batches = batch_segments(
-            filtered_segments,
-            VAD_CONFIG["batch_max_files"],
-            VAD_CONFIG["batch_max_duration"],
-        )
-
-        batch_index = 0
-        total_batches = max(
-            1, int(total_duration / VAD_CONFIG["batch_max_duration"]) + 1
-        )
-
-        for batch in batches:
-            batch_index += 1
-            audio_segments = [seg[2] for seg in batch]
-            results = transcribe_batch(self.model, audio_segments)
-
-            for text, words in emit_results(
-                results,
-                batch,
-                batch_index,
-                total_batches,
-            ):
-                if not text:
-                    continue
-                all_text_parts.append(text)
-                all_words.extend(words)
-
-            processed_duration += sum(len(seg[2]) / float(SAMPLERATE) for seg in batch)
-
-        combined_text = " ".join(all_text_parts)
-        return {"text": combined_text, "words": all_words}
-
-
-@app.function(
-    scaledown_window=60,
-    timeout=600,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
-    image=image,
-)
-@modal.concurrent(max_inputs=40)
-@modal.asgi_app()
-def web():
-    import os
-    import uuid
-
-    from fastapi import (
-        Body,
-        Depends,
-        FastAPI,
-        Form,
-        HTTPException,
-        UploadFile,
-        status,
-    )
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    transcriber_live = TranscriberParakeetLive()
-    transcriber_file = TranscriberParakeetFile()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
-            return
-        raise HTTPException(
-            status_code=status.HTTP_401_UNAUTHORIZED,
-            detail="Invalid API key",
-            headers={"WWW-Authenticate": "Bearer"},
-        )
-
-    class TranscriptResponse(BaseModel):
-        result: dict
-
-    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
-    def transcribe(
-        file: UploadFile = None,
-        files: list[UploadFile] | None = None,
-        model: str = Form(MODEL_NAME),
-        language: str = Form("en"),
-        batch: bool = Form(False),
-    ):
-        # Parakeet only supports English
-        if language != "en":
-            raise HTTPException(
-                status_code=400,
-                detail=f"Parakeet model only supports English. Got language='{language}'",
-            )
-        # Handle both single file and multiple files
-        if not file and not files:
-            raise HTTPException(
-                status_code=400, detail="Either 'file' or 'files' parameter is required"
-            )
-        if batch and not files:
-            raise HTTPException(
-                status_code=400, detail="Batch transcription requires 'files'"
-            )
-
-        upload_files = [file] if file else files
-
-        # Upload files to volume
-        uploaded_filenames = []
-        for upload_file in upload_files:
-            audio_suffix = upload_file.filename.split(".")[-1]
-            assert audio_suffix in SUPPORTED_FILE_EXTENSIONS
-
-            # Generate unique filename
-            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
-            file_path = f"{UPLOADS_PATH}/{unique_filename}"
-
-            print(f"Writing file to: {file_path}")
-            with open(file_path, "wb") as f:
-                content = upload_file.file.read()
-                f.write(content)
-
-            uploaded_filenames.append(unique_filename)
-
-        upload_volume.commit()
-
-        try:
-            # Use A10G live transcriber for per-file transcription
-            if batch and len(upload_files) > 1:
-                # Use batch transcription
-                func = transcriber_live.transcribe_batch.spawn(
-                    filenames=uploaded_filenames,
-                )
-                results = func.get()
-                return {"results": results}
-
-            # Per-file transcription
-            results = []
-            for filename in uploaded_filenames:
-                func = transcriber_live.transcribe_segment.spawn(
-                    filename=filename,
-                )
-                result = func.get()
-                result["filename"] = filename
-                results.append(result)
-
-            return {"results": results} if len(results) > 1 else results[0]
-
-        finally:
-            for filename in uploaded_filenames:
-                try:
-                    file_path = f"{UPLOADS_PATH}/{filename}"
-                    print(f"Deleting file: {file_path}")
-                    os.remove(file_path)
-                except Exception as e:
-                    print(f"Error deleting {filename}: {e}")
-
-            upload_volume.commit()
-
-    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
-    def transcribe_from_url(
-        audio_file_url: str = Body(
-            ..., description="URL of the audio file to transcribe"
-        ),
-        model: str = Body(MODEL_NAME),
-        language: str = Body("en", description="Language code (only 'en' supported)"),
-        timestamp_offset: float = Body(0.0),
-    ):
-        # Parakeet only supports English
-        if language != "en":
-            raise HTTPException(
-                status_code=400,
-                detail=f"Parakeet model only supports English. Got language='{language}'",
-            )
-        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
-
-        try:
-            func = transcriber_file.transcribe_segment.spawn(
-                filename=unique_filename,
-                timestamp_offset=timestamp_offset,
-            )
-            result = func.get()
-            return result
-        finally:
-            try:
-                file_path = f"{UPLOADS_PATH}/{unique_filename}"
-                print(f"Deleting file: {file_path}")
-                os.remove(file_path)
-                upload_volume.commit()
-            except Exception as e:
-                print(f"Error cleaning up {unique_filename}: {e}")
-
-    return app
-
-
-class NoStdStreams:
-    def __init__(self):
-        self.devnull = open(os.devnull, "w")
-
-    def __enter__(self):
-        self._stdout, self._stderr = sys.stdout, sys.stderr
-        self._stdout.flush()
-        self._stderr.flush()
-        sys.stdout, sys.stderr = self.devnull, self.devnull
-
-    def __exit__(self, exc_type, exc_value, traceback):
-        sys.stdout, sys.stderr = self._stdout, self._stderr
-        self.devnull.close()

server/migrations/versions/0194f65cd6d3_add_webhook_fields_to_rooms.py

@@ -0,0 +1,36 @@
+"""Add webhook fields to rooms
+
+Revision ID: 0194f65cd6d3
+Revises: 5a8907fd1d78
+Create Date: 2025-08-27 09:03:19.610995
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0194f65cd6d3"
+down_revision: Union[str, None] = "5a8907fd1d78"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("webhook_url", sa.String(), nullable=True))
+        batch_op.add_column(sa.Column("webhook_secret", sa.String(), nullable=True))
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_column("webhook_secret")
+        batch_op.drop_column("webhook_url")
+
+    # ### end Alembic commands ###

server/migrations/versions/0ce521cda2ee_remove_user_id_from_meeting_table.py

@@ -0,0 +1,36 @@
+"""remove user_id from meeting table
+
+Revision ID: 0ce521cda2ee
+Revises: 6dec9fb5b46c
+Create Date: 2025-09-10 12:40:55.688899
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0ce521cda2ee"
+down_revision: Union[str, None] = "6dec9fb5b46c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("user_id")
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column("user_id", sa.VARCHAR(), autoincrement=False, nullable=True)
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/1e49625677e4_add_platform_support.py

@@ -0,0 +1,50 @@
+"""add_platform_support
+
+Revision ID: 1e49625677e4
+Revises: 9e3f7b2a4c8e
+Create Date: 2025-10-08 13:17:29.943612
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "1e49625677e4"
+down_revision: Union[str, None] = "9e3f7b2a4c8e"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Add platform field with default 'whereby' for backward compatibility."""
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                "platform",
+                sa.String(),
+                nullable=True,
+                server_default=None,
+            )
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column(
+                "platform",
+                sa.String(),
+                nullable=False,
+                server_default="whereby",
+            )
+        )
+
+
+def downgrade() -> None:
+    """Remove platform field."""
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("platform")
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_column("platform")

server/migrations/versions/2ae3db106d4e_clean_up_orphaned_room_id_references_in_.py

@@ -0,0 +1,32 @@
+"""clean up orphaned room_id references in meeting table
+
+Revision ID: 2ae3db106d4e
+Revises: def1b5867d4c
+Create Date: 2025-09-11 10:35:15.759967
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "2ae3db106d4e"
+down_revision: Union[str, None] = "def1b5867d4c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Set room_id to NULL for meetings that reference non-existent rooms
+    op.execute("""
+        UPDATE meeting
+        SET room_id = NULL
+        WHERE room_id IS NOT NULL
+          AND room_id NOT IN (SELECT id FROM room WHERE id IS NOT NULL)
+    """)
+
+
+def downgrade() -> None:
+    # Cannot restore orphaned references - no operation needed
+    pass

server/migrations/versions/2b92a1b03caa_add_daily_participant_session_table_.py

@@ -0,0 +1,79 @@
+"""add daily participant session table with immutable left_at
+
+Revision ID: 2b92a1b03caa
+Revises: f8294b31f022
+Create Date: 2025-11-13 20:29:30.486577
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "2b92a1b03caa"
+down_revision: Union[str, None] = "f8294b31f022"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Create table
+    op.create_table(
+        "daily_participant_session",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("meeting_id", sa.String(), nullable=False),
+        sa.Column("room_id", sa.String(), nullable=False),
+        sa.Column("session_id", sa.String(), nullable=False),
+        sa.Column("user_id", sa.String(), nullable=True),
+        sa.Column("user_name", sa.String(), nullable=False),
+        sa.Column("joined_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("left_at", sa.DateTime(timezone=True), nullable=True),
+        sa.ForeignKeyConstraint(["meeting_id"], ["meeting.id"], ondelete="CASCADE"),
+        sa.ForeignKeyConstraint(["room_id"], ["room.id"], ondelete="CASCADE"),
+        sa.PrimaryKeyConstraint("id"),
+    )
+    with op.batch_alter_table("daily_participant_session", schema=None) as batch_op:
+        batch_op.create_index(
+            "idx_daily_session_meeting_left", ["meeting_id", "left_at"], unique=False
+        )
+        batch_op.create_index("idx_daily_session_room", ["room_id"], unique=False)
+
+    # Create trigger function to prevent left_at from being updated once set
+    op.execute("""
+        CREATE OR REPLACE FUNCTION prevent_left_at_update()
+        RETURNS TRIGGER AS $$
+        BEGIN
+            IF OLD.left_at IS NOT NULL THEN
+                RAISE EXCEPTION 'left_at is immutable once set';
+            END IF;
+            RETURN NEW;
+        END;
+        $$ LANGUAGE plpgsql;
+    """)
+
+    # Create trigger
+    op.execute("""
+        CREATE TRIGGER prevent_left_at_update_trigger
+        BEFORE UPDATE ON daily_participant_session
+        FOR EACH ROW
+        EXECUTE FUNCTION prevent_left_at_update();
+    """)
+
+
+def downgrade() -> None:
+    # Drop trigger
+    op.execute(
+        "DROP TRIGGER IF EXISTS prevent_left_at_update_trigger ON daily_participant_session;"
+    )
+
+    # Drop trigger function
+    op.execute("DROP FUNCTION IF EXISTS prevent_left_at_update();")
+
+    # Drop indexes and table
+    with op.batch_alter_table("daily_participant_session", schema=None) as batch_op:
+        batch_op.drop_index("idx_daily_session_room")
+        batch_op.drop_index("idx_daily_session_meeting_left")
+
+    op.drop_table("daily_participant_session")

server/migrations/versions/5a8907fd1d78_add_cascade_delete_to_meeting_consent_.py

@@ -0,0 +1,50 @@
+"""add cascade delete to meeting consent foreign key
+
+Revision ID: 5a8907fd1d78
+Revises: 0ab2d7ffaa16
+Create Date: 2025-08-26 17:26:50.945491
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "5a8907fd1d78"
+down_revision: Union[str, None] = "0ab2d7ffaa16"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+            ondelete="CASCADE",
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/5d6b9df9b045_make_room_platform_non_nullable_with_.py

@@ -0,0 +1,30 @@
+"""Make room platform non-nullable with dynamic default
+
+Revision ID: 5d6b9df9b045
+Revises: 2b92a1b03caa
+Create Date: 2025-11-21 13:22:25.756584
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "5d6b9df9b045"
+down_revision: Union[str, None] = "2b92a1b03caa"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.execute("UPDATE room SET platform = 'whereby' WHERE platform IS NULL")
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column("platform", existing_type=sa.String(), nullable=False)
+
+
+def downgrade() -> None:
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column("platform", existing_type=sa.String(), nullable=True)

server/migrations/versions/6025e9b2bef2_remove_one_active_meeting_per_room_.py

@@ -0,0 +1,53 @@
+"""remove_one_active_meeting_per_room_constraint
+
+Revision ID: 6025e9b2bef2
+Revises: 2ae3db106d4e
+Create Date: 2025-08-18 18:45:44.418392
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6025e9b2bef2"
+down_revision: Union[str, None] = "2ae3db106d4e"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove the unique constraint that prevents multiple active meetings per room
+    # This is needed to support calendar integration with overlapping meetings
+    # Check if index exists before trying to drop it
+    from alembic import context
+
+    if context.get_context().dialect.name == "postgresql":
+        conn = op.get_bind()
+        result = conn.execute(
+            sa.text(
+                "SELECT 1 FROM pg_indexes WHERE indexname = 'idx_one_active_meeting_per_room'"
+            )
+        )
+        if result.fetchone():
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+    else:
+        # For SQLite, just try to drop it
+        try:
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+        except:
+            pass
+
+
+def downgrade() -> None:
+    # Restore the unique constraint
+    op.create_index(
+        "idx_one_active_meeting_per_room",
+        "meeting",
+        ["room_id"],
+        unique=True,
+        postgresql_where=sa.text("is_active = true"),
+        sqlite_where=sa.text("is_active = 1"),
+    )

server/migrations/versions/61882a919591_webhook_url_and_secret_null_by_default.py

@@ -0,0 +1,28 @@
+"""webhook url and secret null by default
+
+
+Revision ID: 61882a919591
+Revises: 0194f65cd6d3
+Create Date: 2025-08-29 11:46:36.738091
+
+"""
+
+from typing import Sequence, Union
+
+# revision identifiers, used by Alembic.
+revision: str = "61882a919591"
+down_revision: Union[str, None] = "0194f65cd6d3"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###

server/migrations/versions/6dec9fb5b46c_make_meeting_room_id_required_and_add_.py

@@ -0,0 +1,35 @@
+"""make meeting room_id required and add foreign key
+
+Revision ID: 6dec9fb5b46c
+Revises: 61882a919591
+Create Date: 2025-09-10 10:47:06.006819
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6dec9fb5b46c"
+down_revision: Union[str, None] = "61882a919591"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.create_foreign_key(
+            None, "room", ["room_id"], ["id"], ondelete="CASCADE"
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("meeting_room_id_fkey", type_="foreignkey")
+
+    # ### end Alembic commands ###

server/migrations/versions/9e3f7b2a4c8e_add_user_api_keys.py

@@ -0,0 +1,38 @@
+"""add user api keys
+
+Revision ID: 9e3f7b2a4c8e
+Revises: dc035ff72fd5
+Create Date: 2025-10-17 00:00:00.000000
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "9e3f7b2a4c8e"
+down_revision: Union[str, None] = "dc035ff72fd5"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "user_api_key",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("user_id", sa.String(), nullable=False),
+        sa.Column("key_hash", sa.String(), nullable=False),
+        sa.Column("name", sa.String(), nullable=True),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.PrimaryKeyConstraint("id"),
+    )
+
+    with op.batch_alter_table("user_api_key", schema=None) as batch_op:
+        batch_op.create_index("idx_user_api_key_hash", ["key_hash"], unique=True)
+        batch_op.create_index("idx_user_api_key_user_id", ["user_id"], unique=False)
+
+
+def downgrade() -> None:
+    op.drop_table("user_api_key")

server/migrations/versions/bbafedfa510c_add_user_table.py

@@ -0,0 +1,38 @@
+"""add user table
+
+Revision ID: bbafedfa510c
+Revises: 5d6b9df9b045
+Create Date: 2025-11-19 21:06:30.543262
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "bbafedfa510c"
+down_revision: Union[str, None] = "5d6b9df9b045"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "user",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("email", sa.String(), nullable=False),
+        sa.Column("authentik_uid", sa.String(), nullable=False),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+        sa.PrimaryKeyConstraint("id"),
+    )
+
+    with op.batch_alter_table("user", schema=None) as batch_op:
+        batch_op.create_index("idx_user_authentik_uid", ["authentik_uid"], unique=True)
+        batch_op.create_index("idx_user_email", ["email"], unique=False)
+
+
+def downgrade() -> None:
+    op.drop_table("user")

server/migrations/versions/d4a1c446458c_add_grace_period_fields_to_meeting.py

@@ -0,0 +1,34 @@
+"""add_grace_period_fields_to_meeting
+
+Revision ID: d4a1c446458c
+Revises: 6025e9b2bef2
+Create Date: 2025-08-18 18:50:37.768052
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "d4a1c446458c"
+down_revision: Union[str, None] = "6025e9b2bef2"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Add fields to track when participants left for grace period logic
+    op.add_column(
+        "meeting", sa.Column("last_participant_left_at", sa.DateTime(timezone=True))
+    )
+    op.add_column(
+        "meeting",
+        sa.Column("grace_period_minutes", sa.Integer, server_default=sa.text("15")),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("meeting", "grace_period_minutes")
+    op.drop_column("meeting", "last_participant_left_at")

server/migrations/versions/d8e204bbf615_add_calendar.py

@@ -0,0 +1,129 @@
+"""add calendar
+
+Revision ID: d8e204bbf615
+Revises: d4a1c446458c
+Create Date: 2025-09-10 19:56:22.295756
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "d8e204bbf615"
+down_revision: Union[str, None] = "d4a1c446458c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "calendar_event",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("room_id", sa.String(), nullable=False),
+        sa.Column("ics_uid", sa.Text(), nullable=False),
+        sa.Column("title", sa.Text(), nullable=True),
+        sa.Column("description", sa.Text(), nullable=True),
+        sa.Column("start_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("end_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("attendees", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+        sa.Column("location", sa.Text(), nullable=True),
+        sa.Column("ics_raw_data", sa.Text(), nullable=True),
+        sa.Column("last_synced", sa.DateTime(timezone=True), nullable=False),
+        sa.Column(
+            "is_deleted", sa.Boolean(), server_default=sa.text("false"), nullable=False
+        ),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["room_id"],
+            ["room.id"],
+            name="fk_calendar_event_room_id",
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint("room_id", "ics_uid", name="uq_room_calendar_event"),
+    )
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.create_index(
+            "idx_calendar_event_deleted",
+            ["is_deleted"],
+            unique=False,
+            postgresql_where=sa.text("NOT is_deleted"),
+        )
+        batch_op.create_index(
+            "idx_calendar_event_room_start", ["room_id", "start_time"], unique=False
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("calendar_event_id", sa.String(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "calendar_metadata",
+                postgresql.JSONB(astext_type=sa.Text()),
+                nullable=True,
+            )
+        )
+        batch_op.create_index(
+            "idx_meeting_calendar_event", ["calendar_event_id"], unique=False
+        )
+        batch_op.create_foreign_key(
+            "fk_meeting_calendar_event_id",
+            "calendar_event",
+            ["calendar_event_id"],
+            ["id"],
+            ondelete="SET NULL",
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("ics_url", sa.Text(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "ics_fetch_interval", sa.Integer(), server_default="300", nullable=True
+            )
+        )
+        batch_op.add_column(
+            sa.Column(
+                "ics_enabled",
+                sa.Boolean(),
+                server_default=sa.text("false"),
+                nullable=False,
+            )
+        )
+        batch_op.add_column(
+            sa.Column("ics_last_sync", sa.DateTime(timezone=True), nullable=True)
+        )
+        batch_op.add_column(sa.Column("ics_last_etag", sa.Text(), nullable=True))
+        batch_op.create_index("idx_room_ics_enabled", ["ics_enabled"], unique=False)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_index("idx_room_ics_enabled")
+        batch_op.drop_column("ics_last_etag")
+        batch_op.drop_column("ics_last_sync")
+        batch_op.drop_column("ics_enabled")
+        batch_op.drop_column("ics_fetch_interval")
+        batch_op.drop_column("ics_url")
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("fk_meeting_calendar_event_id", type_="foreignkey")
+        batch_op.drop_index("idx_meeting_calendar_event")
+        batch_op.drop_column("calendar_metadata")
+        batch_op.drop_column("calendar_event_id")
+
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.drop_index("idx_calendar_event_room_start")
+        batch_op.drop_index(
+            "idx_calendar_event_deleted", postgresql_where=sa.text("NOT is_deleted")
+        )
+
+    op.drop_table("calendar_event")
+    # ### end Alembic commands ###

server/migrations/versions/dc035ff72fd5_remove_grace_period_fields.py

@@ -0,0 +1,43 @@
+"""remove_grace_period_fields
+
+Revision ID: dc035ff72fd5
+Revises: d8e204bbf615
+Create Date: 2025-09-11 10:36:45.197588
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "dc035ff72fd5"
+down_revision: Union[str, None] = "d8e204bbf615"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove grace period columns from meeting table
+    op.drop_column("meeting", "last_participant_left_at")
+    op.drop_column("meeting", "grace_period_minutes")
+
+
+def downgrade() -> None:
+    # Add back grace period columns to meeting table
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "last_participant_left_at", sa.DateTime(timezone=True), nullable=True
+        ),
+    )
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "grace_period_minutes",
+            sa.Integer(),
+            server_default=sa.text("15"),
+            nullable=True,
+        ),
+    )

server/migrations/versions/def1b5867d4c_make_meeting_room_id_nullable_but_keep_.py

@@ -0,0 +1,34 @@
+"""make meeting room_id nullable but keep foreign key
+
+Revision ID: def1b5867d4c
+Revises: 0ce521cda2ee
+Create Date: 2025-09-11 09:42:18.697264
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "def1b5867d4c"
+down_revision: Union[str, None] = "0ce521cda2ee"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=True)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=False)
+
+    # ### end Alembic commands ###

server/migrations/versions/f8294b31f022_add_track_keys.py

@@ -0,0 +1,28 @@
+"""add_track_keys
+
+Revision ID: f8294b31f022
+Revises: 1e49625677e4
+Create Date: 2025-10-27 18:52:17.589167
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "f8294b31f022"
+down_revision: Union[str, None] = "1e49625677e4"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("track_keys", sa.JSON(), nullable=True))
+
+
+def downgrade() -> None:
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.drop_column("track_keys")

server/pyproject.toml

@@ -12,7 +12,6 @@ dependencies = [
     "requests>=2.31.0",
     "aiortc>=1.5.0",
     "sortedcontainers>=2.4.0",
-    "loguru>=0.7.0",
     "pydantic-settings>=2.0.2",
     "structlog>=23.1.0",
     "uvicorn[standard]>=0.23.1",
@@ -27,7 +26,6 @@ dependencies = [
     "prometheus-fastapi-instrumentator>=6.1.0",
     "sentencepiece>=0.1.99",
     "protobuf>=4.24.3",
-    "profanityfilter>=2.0.6",
     "celery>=5.3.4",
     "redis>=5.0.1",
     "python-jose[cryptography]>=3.3.0",
@@ -40,6 +38,7 @@ dependencies = [
     "llama-index-llms-openai-like>=0.4.0",
     "pytest-env>=1.1.5",
     "webvtt-py>=0.5.0",
+    "icalendar>=6.0.0",
 ]
 
 [dependency-groups]
@@ -113,13 +112,14 @@ source = ["reflector"]
 [tool.pytest_env]
 ENVIRONMENT = "pytest"
 DATABASE_URL = "postgresql://test_user:test_password@localhost:15432/reflector_test"
+AUTH_BACKEND = "jwt"
 
 [tool.pytest.ini_options]
 addopts = "-ra -q --disable-pytest-warnings --cov --cov-report html -v"
 testpaths = ["tests"]
 asyncio_mode = "auto"
 markers = [
-    "gpu_modal: mark test to run only with GPU Modal endpoints (deselect with '-m \"not gpu_modal\"')",
+    "model_api: tests for the unified model-serving HTTP API (backend- and hardware-agnostic)",
 ]
 
 [tool.ruff.lint]
@@ -131,7 +131,7 @@ select = [
 
 [tool.ruff.lint.per-file-ignores]
 "reflector/processors/summary/summary_builder.py" = ["E501"]
-"gpu/**.py" = ["PLC0415"]
+"gpu/modal_deployments/**.py" = ["PLC0415"]
 "reflector/tools/**.py" = ["PLC0415"]
 "migrations/versions/**.py" = ["PLC0415"]
 "tests/**.py" = ["PLC0415"]

server/reflector/app.py

@@ -12,6 +12,7 @@ from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.logger import logger
 from reflector.metrics import metrics_init
 from reflector.settings import settings
+from reflector.views.daily import router as daily_router
 from reflector.views.meetings import router as meetings_router
 from reflector.views.rooms import router as rooms_router
 from reflector.views.rtc_offer import router as rtc_offer_router
@@ -26,6 +27,8 @@ from reflector.views.transcripts_upload import router as transcripts_upload_rout
 from reflector.views.transcripts_webrtc import router as transcripts_webrtc_router
 from reflector.views.transcripts_websocket import router as transcripts_websocket_router
 from reflector.views.user import router as user_router
+from reflector.views.user_api_keys import router as user_api_keys_router
+from reflector.views.user_websocket import router as user_ws_router
 from reflector.views.whereby import router as whereby_router
 from reflector.views.zulip import router as zulip_router
 
@@ -65,6 +68,12 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+
+@app.get("/health")
+async def health():
+    return {"status": "healthy"}
+
+
 # metrics
 instrumentator = Instrumentator(
     excluded_handlers=["/docs", "/metrics"],
@@ -84,8 +93,11 @@ app.include_router(transcripts_websocket_router, prefix="/v1")
 app.include_router(transcripts_webrtc_router, prefix="/v1")
 app.include_router(transcripts_process_router, prefix="/v1")
 app.include_router(user_router, prefix="/v1")
+app.include_router(user_api_keys_router, prefix="/v1")
+app.include_router(user_ws_router, prefix="/v1")
 app.include_router(zulip_router, prefix="/v1")
 app.include_router(whereby_router, prefix="/v1")
+app.include_router(daily_router, prefix="/v1/daily")
 add_pagination(app)
 
 # prepare celery

server/reflector/asynctask.py

@@ -0,0 +1,27 @@
+import asyncio
+import functools
+
+from reflector.db import get_database
+
+
+def asynctask(f):
+    @functools.wraps(f)
+    def wrapper(*args, **kwargs):
+        async def run_with_db():
+            database = get_database()
+            await database.connect()
+            try:
+                return await f(*args, **kwargs)
+            finally:
+                await database.disconnect()
+
+        coro = run_with_db()
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            loop = None
+        if loop and loop.is_running():
+            return loop.run_until_complete(coro)
+        return asyncio.run(coro)
+
+    return wrapper

server/reflector/auth/auth_jwt.py

@@ -1,14 +1,18 @@
-from typing import Annotated, Optional
+from typing import Annotated, List, Optional
 
 from fastapi import Depends, HTTPException
-from fastapi.security import OAuth2PasswordBearer
+from fastapi.security import APIKeyHeader, OAuth2PasswordBearer
 from jose import JWTError, jwt
 from pydantic import BaseModel
 
+from reflector.db.user_api_keys import user_api_keys_controller
+from reflector.db.users import user_controller
 from reflector.logger import logger
 from reflector.settings import settings
+from reflector.utils import generate_uuid4
 
 oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token", auto_error=False)
+api_key_header = APIKeyHeader(name="X-API-Key", auto_error=False)
 
 jwt_public_key = open(f"reflector/auth/jwt/keys/{settings.AUTH_JWT_PUBLIC_KEY}").read()
 jwt_algorithm = settings.AUTH_JWT_ALGORITHM
@@ -26,7 +30,7 @@ class JWTException(Exception):
 
 class UserInfo(BaseModel):
     sub: str
-    email: str
+    email: Optional[str] = None
 
     def __getitem__(self, key):
         return getattr(self, key)
@@ -58,33 +62,65 @@ def authenticated(token: Annotated[str, Depends(oauth2_scheme)]):
     return None
 
 
-def current_user(
-    token: Annotated[Optional[str], Depends(oauth2_scheme)],
-    jwtauth: JWTAuth = Depends(),
-):
-    if token is None:
-        raise HTTPException(status_code=401, detail="Not authenticated")
-    try:
-        payload = jwtauth.verify_token(token)
-        sub = payload["sub"]
-        return UserInfo(sub=sub)
-    except JWTError as e:
-        logger.error(f"JWT error: {e}")
-        raise HTTPException(status_code=401, detail="Invalid authentication")
+async def _authenticate_user(
+    jwt_token: Optional[str],
+    api_key: Optional[str],
+    jwtauth: JWTAuth,
+) -> UserInfo | None:
+    user_infos: List[UserInfo] = []
+    if api_key:
+        user_api_key = await user_api_keys_controller.verify_key(api_key)
+        if user_api_key:
+            user_infos.append(UserInfo(sub=user_api_key.user_id, email=None))
 
+    if jwt_token:
+        try:
+            payload = jwtauth.verify_token(jwt_token)
+            authentik_uid = payload["sub"]
+            email = payload["email"]
 
-def current_user_optional(
-    token: Annotated[Optional[str], Depends(oauth2_scheme)],
-    jwtauth: JWTAuth = Depends(),
-):
-    # we accept no token, but if one is provided, it must be a valid one.
-    if token is None:
+            user = await user_controller.get_by_authentik_uid(authentik_uid)
+            if not user:
+                logger.info(
+                    f"Creating new user on first login: {authentik_uid} ({email})"
+                )
+                user = await user_controller.create_or_update(
+                    id=generate_uuid4(),
+                    authentik_uid=authentik_uid,
+                    email=email,
+                )
+
+            user_infos.append(UserInfo(sub=user.id, email=email))
+        except JWTError as e:
+            logger.error(f"JWT error: {e}")
+            raise HTTPException(status_code=401, detail="Invalid authentication")
+
+    if len(user_infos) == 0:
         return None
-    try:
-        payload = jwtauth.verify_token(token)
-        sub = payload["sub"]
-        email = payload["email"]
-        return UserInfo(sub=sub, email=email)
-    except JWTError as e:
-        logger.error(f"JWT error: {e}")
-        raise HTTPException(status_code=401, detail="Invalid authentication")
+
+    if len(set([x.sub for x in user_infos])) > 1:
+        raise JWTException(
+            status_code=401,
+            detail="Invalid authentication: more than one user provided",
+        )
+
+    return user_infos[0]
+
+
+async def current_user(
+    jwt_token: Annotated[Optional[str], Depends(oauth2_scheme)],
+    api_key: Annotated[Optional[str], Depends(api_key_header)],
+    jwtauth: JWTAuth = Depends(),
+):
+    user = await _authenticate_user(jwt_token, api_key, jwtauth)
+    if user is None:
+        raise HTTPException(status_code=401, detail="Not authenticated")
+    return user
+
+
+async def current_user_optional(
+    jwt_token: Annotated[Optional[str], Depends(oauth2_scheme)],
+    api_key: Annotated[Optional[str], Depends(api_key_header)],
+    jwtauth: JWTAuth = Depends(),
+):
+    return await _authenticate_user(jwt_token, api_key, jwtauth)

server/reflector/dailyco_api/README.md

@@ -0,0 +1,6 @@
+anything about Daily.co api interaction
+
+- webhook event shapes
+- REST api client
+
+No REST api client existing found in the wild; the official lib is about working with videocall as a bot

server/reflector/dailyco_api/__init__.py

@@ -0,0 +1,108 @@
+"""
+Daily.co API Module
+"""
+
+# Client
+from .client import DailyApiClient, DailyApiError
+
+# Request models
+from .requests import (
+    CreateMeetingTokenRequest,
+    CreateRoomRequest,
+    CreateWebhookRequest,
+    MeetingTokenProperties,
+    RecordingsBucketConfig,
+    RoomProperties,
+    UpdateWebhookRequest,
+)
+
+# Response models
+from .responses import (
+    MeetingParticipant,
+    MeetingParticipantsResponse,
+    MeetingResponse,
+    MeetingTokenResponse,
+    RecordingResponse,
+    RecordingS3Info,
+    RoomPresenceParticipant,
+    RoomPresenceResponse,
+    RoomResponse,
+    WebhookResponse,
+)
+
+# Webhook utilities
+from .webhook_utils import (
+    extract_room_name,
+    parse_participant_joined,
+    parse_participant_left,
+    parse_recording_error,
+    parse_recording_ready,
+    parse_recording_started,
+    parse_webhook_payload,
+    verify_webhook_signature,
+)
+
+# Webhook models
+from .webhooks import (
+    DailyTrack,
+    DailyWebhookEvent,
+    DailyWebhookEventUnion,
+    ParticipantJoinedEvent,
+    ParticipantJoinedPayload,
+    ParticipantLeftEvent,
+    ParticipantLeftPayload,
+    RecordingErrorEvent,
+    RecordingErrorPayload,
+    RecordingReadyEvent,
+    RecordingReadyToDownloadPayload,
+    RecordingStartedEvent,
+    RecordingStartedPayload,
+)
+
+__all__ = [
+    # Client
+    "DailyApiClient",
+    "DailyApiError",
+    # Requests
+    "CreateRoomRequest",
+    "RoomProperties",
+    "RecordingsBucketConfig",
+    "CreateMeetingTokenRequest",
+    "MeetingTokenProperties",
+    "CreateWebhookRequest",
+    "UpdateWebhookRequest",
+    # Responses
+    "RoomResponse",
+    "RoomPresenceResponse",
+    "RoomPresenceParticipant",
+    "MeetingParticipantsResponse",
+    "MeetingParticipant",
+    "MeetingResponse",
+    "RecordingResponse",
+    "RecordingS3Info",
+    "MeetingTokenResponse",
+    "WebhookResponse",
+    # Webhooks
+    "DailyWebhookEvent",
+    "DailyWebhookEventUnion",
+    "DailyTrack",
+    "ParticipantJoinedEvent",
+    "ParticipantJoinedPayload",
+    "ParticipantLeftEvent",
+    "ParticipantLeftPayload",
+    "RecordingStartedEvent",
+    "RecordingStartedPayload",
+    "RecordingReadyEvent",
+    "RecordingReadyToDownloadPayload",
+    "RecordingErrorEvent",
+    "RecordingErrorPayload",
+    # Webhook utilities
+    "verify_webhook_signature",
+    "extract_room_name",
+    "parse_webhook_payload",
+    "parse_participant_joined",
+    "parse_participant_left",
+    "parse_recording_started",
+    "parse_recording_ready",
+    "parse_recording_error",
+]

server/reflector/dailyco_api/client.py

@@ -0,0 +1,573 @@
+"""
+Daily.co API Client
+
+Complete async client for Daily.co REST API with Pydantic models.
+
+Reference: https://docs.daily.co/reference/rest-api
+"""
+
+from http import HTTPStatus
+from typing import Any
+
+import httpx
+import structlog
+
+from reflector.utils.string import NonEmptyString
+
+from .requests import (
+    CreateMeetingTokenRequest,
+    CreateRoomRequest,
+    CreateWebhookRequest,
+    UpdateWebhookRequest,
+)
+from .responses import (
+    MeetingParticipantsResponse,
+    MeetingResponse,
+    MeetingTokenResponse,
+    RecordingResponse,
+    RoomPresenceResponse,
+    RoomResponse,
+    WebhookResponse,
+)
+
+logger = structlog.get_logger(__name__)
+
+
+class DailyApiError(Exception):
+    """Daily.co API error with full request/response context."""
+
+    def __init__(self, operation: str, response: httpx.Response):
+        self.operation = operation
+        self.response = response
+        self.status_code = response.status_code
+        self.response_body = response.text
+        self.url = str(response.url)
+        self.request_body = (
+            response.request.content.decode() if response.request.content else None
+        )
+
+        super().__init__(
+            f"Daily.co API error: {operation} failed with status {self.status_code}"
+        )
+
+
+class DailyApiClient:
+    """
+    Complete async client for Daily.co REST API.
+
+    Usage:
+        # Direct usage
+        client = DailyApiClient(api_key="your_api_key")
+        room = await client.create_room(CreateRoomRequest(name="my-room"))
+        await client.close()  # Clean up when done
+
+        # Context manager (recommended)
+        async with DailyApiClient(api_key="your_api_key") as client:
+            room = await client.create_room(CreateRoomRequest(name="my-room"))
+    """
+
+    BASE_URL = "https://api.daily.co/v1"
+    DEFAULT_TIMEOUT = 10.0
+
+    def __init__(
+        self,
+        api_key: NonEmptyString,
+        webhook_secret: NonEmptyString | None = None,
+        timeout: float = DEFAULT_TIMEOUT,
+        base_url: NonEmptyString | None = None,
+    ):
+        """
+        Initialize Daily.co API client.
+
+        Args:
+            api_key: Daily.co API key (Bearer token)
+            webhook_secret: Base64-encoded HMAC secret for webhook verification.
+                Must match the 'hmac' value provided when creating webhooks.
+                Generate with: base64.b64encode(os.urandom(32)).decode()
+            timeout: Default request timeout in seconds
+            base_url: Override base URL (for testing)
+        """
+        self.api_key = api_key
+        self.webhook_secret = webhook_secret
+        self.timeout = timeout
+        self.base_url = base_url or self.BASE_URL
+
+        self.headers = {
+            "Authorization": f"Bearer {api_key}",
+            "Content-Type": "application/json",
+        }
+
+        self._client: httpx.AsyncClient | None = None
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        await self.close()
+
+    async def _get_client(self) -> httpx.AsyncClient:
+        if self._client is None:
+            self._client = httpx.AsyncClient(timeout=self.timeout)
+        return self._client
+
+    async def close(self):
+        if self._client is not None:
+            await self._client.aclose()
+            self._client = None
+
+    async def _handle_response(
+        self, response: httpx.Response, operation: str
+    ) -> dict[str, Any]:
+        """
+        Handle API response with error logging.
+
+        Args:
+            response: HTTP response
+            operation: Operation name for logging (e.g., "create_room")
+
+        Returns:
+            Parsed JSON response
+
+        Raises:
+            DailyApiError: If request failed with full context
+        """
+        if response.status_code >= 400:
+            logger.error(
+                f"Daily.co API error: {operation}",
+                status_code=response.status_code,
+                response_body=response.text,
+                request_body=response.request.content.decode()
+                if response.request.content
+                else None,
+                url=str(response.url),
+            )
+            raise DailyApiError(operation, response)
+
+        return response.json()
+
+    # ============================================================================
+    # ROOMS
+    # ============================================================================
+
+    async def create_room(self, request: CreateRoomRequest) -> RoomResponse:
+        """
+        Create a new Daily.co room.
+
+        Reference: https://docs.daily.co/reference/rest-api/rooms/create-room
+
+        Args:
+            request: Room creation request with name, privacy, and properties
+
+        Returns:
+            Created room data including URL and ID
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.post(
+            f"{self.base_url}/rooms",
+            headers=self.headers,
+            json=request.model_dump(exclude_none=True),
+        )
+
+        data = await self._handle_response(response, "create_room")
+        return RoomResponse(**data)
+
+    async def get_room(self, room_name: NonEmptyString) -> RoomResponse:
+        """
+        Get room configuration.
+
+        Args:
+            room_name: Daily.co room name
+
+        Returns:
+            Room configuration data
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/rooms/{room_name}",
+            headers=self.headers,
+        )
+
+        data = await self._handle_response(response, "get_room")
+        return RoomResponse(**data)
+
+    async def get_room_presence(
+        self, room_name: NonEmptyString
+    ) -> RoomPresenceResponse:
+        """
+        Get current participants in a room (real-time presence).
+
+        Reference: https://docs.daily.co/reference/rest-api/rooms/get-room-presence
+
+        Args:
+            room_name: Daily.co room name
+
+        Returns:
+            List of currently present participants with join time and duration
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/rooms/{room_name}/presence",
+            headers=self.headers,
+        )
+
+        data = await self._handle_response(response, "get_room_presence")
+        return RoomPresenceResponse(**data)
+
+    async def delete_room(self, room_name: NonEmptyString) -> None:
+        """
+        Delete a room (idempotent - succeeds even if room doesn't exist).
+
+        Reference: https://docs.daily.co/reference/rest-api/rooms/delete-room
+
+        Args:
+            room_name: Daily.co room name
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails (except 404)
+        """
+        client = await self._get_client()
+        response = await client.delete(
+            f"{self.base_url}/rooms/{room_name}",
+            headers=self.headers,
+        )
+
+        # Idempotent delete - 404 means already deleted
+        if response.status_code == HTTPStatus.NOT_FOUND:
+            logger.debug("Room not found (already deleted)", room_name=room_name)
+            return
+
+        await self._handle_response(response, "delete_room")
+
+    # ============================================================================
+    # MEETINGS
+    # ============================================================================
+
+    async def get_meeting(self, meeting_id: NonEmptyString) -> MeetingResponse:
+        """
+        Get full meeting information including participants.
+
+        Reference: https://docs.daily.co/reference/rest-api/meetings/get-meeting-information
+
+        Args:
+            meeting_id: Daily.co meeting/session ID
+
+        Returns:
+            Meeting metadata including room, duration, participants, and status
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/meetings/{meeting_id}",
+            headers=self.headers,
+        )
+
+        data = await self._handle_response(response, "get_meeting")
+        return MeetingResponse(**data)
+
+    async def get_meeting_participants(
+        self,
+        meeting_id: NonEmptyString,
+        limit: int | None = None,
+        joined_after: NonEmptyString | None = None,
+        joined_before: NonEmptyString | None = None,
+    ) -> MeetingParticipantsResponse:
+        """
+        Get historical participant data from a completed meeting (paginated).
+
+        Reference: https://docs.daily.co/reference/rest-api/meetings/get-meeting-participants
+
+        Args:
+            meeting_id: Daily.co meeting/session ID
+            limit: Maximum number of participant records to return
+            joined_after: Return participants who joined after this participant_id
+            joined_before: Return participants who joined before this participant_id
+
+        Returns:
+            List of participants with join times and duration
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails (404 when no more participants)
+
+        Note:
+            For pagination, use joined_after with the last participant_id from previous response.
+            Returns 404 when no more participants remain.
+        """
+        params = {}
+        if limit is not None:
+            params["limit"] = limit
+        if joined_after is not None:
+            params["joined_after"] = joined_after
+        if joined_before is not None:
+            params["joined_before"] = joined_before
+
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/meetings/{meeting_id}/participants",
+            headers=self.headers,
+            params=params,
+        )
+
+        data = await self._handle_response(response, "get_meeting_participants")
+        return MeetingParticipantsResponse(**data)
+
+    # ============================================================================
+    # RECORDINGS
+    # ============================================================================
+
+    async def get_recording(self, recording_id: NonEmptyString) -> RecordingResponse:
+        """
+        https://docs.daily.co/reference/rest-api/recordings/get-recording-information
+        Get recording metadata and status.
+        """
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/recordings/{recording_id}",
+            headers=self.headers,
+        )
+
+        data = await self._handle_response(response, "get_recording")
+        return RecordingResponse(**data)
+
+    async def list_recordings(
+        self,
+        room_name: NonEmptyString | None = None,
+        starting_after: str | None = None,
+        ending_before: str | None = None,
+        limit: int = 100,
+    ) -> list[RecordingResponse]:
+        """
+        List recordings with optional filters.
+
+        Reference: https://docs.daily.co/reference/rest-api/recordings
+
+        Args:
+            room_name: Filter by room name
+            starting_after: Pagination cursor - recording ID to start after
+            ending_before: Pagination cursor - recording ID to end before
+            limit: Max results per page (default 100, max 100)
+
+        Note: starting_after/ending_before are pagination cursors (recording IDs),
+        NOT time filters. API returns recordings in reverse chronological order.
+        """
+        client = await self._get_client()
+
+        params = {"limit": limit}
+        if room_name:
+            params["room_name"] = room_name
+        if starting_after:
+            params["starting_after"] = starting_after
+        if ending_before:
+            params["ending_before"] = ending_before
+
+        response = await client.get(
+            f"{self.base_url}/recordings",
+            headers=self.headers,
+            params=params,
+        )
+
+        data = await self._handle_response(response, "list_recordings")
+
+        if not isinstance(data, dict) or "data" not in data:
+            logger.error(
+                "Daily.co API returned unexpected format for list_recordings",
+                data_type=type(data).__name__,
+                data_keys=list(data.keys()) if isinstance(data, dict) else None,
+                data_sample=str(data)[:500],
+                room_name=room_name,
+                operation="list_recordings",
+            )
+            raise httpx.HTTPStatusError(
+                message=f"Unexpected response format from list_recordings: {type(data).__name__}",
+                request=response.request,
+                response=response,
+            )
+
+        return [RecordingResponse(**r) for r in data["data"]]
+
+    # ============================================================================
+    # MEETING TOKENS
+    # ============================================================================
+
+    async def create_meeting_token(
+        self, request: CreateMeetingTokenRequest
+    ) -> MeetingTokenResponse:
+        """
+        Create a meeting token for participant authentication.
+
+        Reference: https://docs.daily.co/reference/rest-api/meeting-tokens/create-meeting-token
+
+        Args:
+            request: Token properties including room name, user_id, permissions
+
+        Returns:
+            JWT meeting token
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.post(
+            f"{self.base_url}/meeting-tokens",
+            headers=self.headers,
+            json=request.model_dump(exclude_none=True),
+        )
+
+        data = await self._handle_response(response, "create_meeting_token")
+        return MeetingTokenResponse(**data)
+
+    # ============================================================================
+    # WEBHOOKS
+    # ============================================================================
+
+    async def list_webhooks(self) -> list[WebhookResponse]:
+        """
+        List all configured webhooks for this account.
+
+        Reference: https://docs.daily.co/reference/rest-api/webhooks
+
+        Returns:
+            List of webhook configurations
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.get(
+            f"{self.base_url}/webhooks",
+            headers=self.headers,
+        )
+
+        data = await self._handle_response(response, "list_webhooks")
+
+        # Daily.co returns array directly (not paginated)
+        if isinstance(data, list):
+            return [WebhookResponse(**wh) for wh in data]
+
+        # Future-proof: handle potential pagination envelope
+        if isinstance(data, dict) and "data" in data:
+            return [WebhookResponse(**wh) for wh in data["data"]]
+
+        logger.warning("Unexpected webhook list response format", data=data)
+        return []
+
+    async def create_webhook(self, request: CreateWebhookRequest) -> WebhookResponse:
+        """
+        Create a new webhook subscription.
+
+        Reference: https://docs.daily.co/reference/rest-api/webhooks
+
+        Args:
+            request: Webhook configuration with URL, event types, and HMAC secret
+
+        Returns:
+            Created webhook with UUID and state
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.post(
+            f"{self.base_url}/webhooks",
+            headers=self.headers,
+            json=request.model_dump(exclude_none=True),
+        )
+
+        data = await self._handle_response(response, "create_webhook")
+        return WebhookResponse(**data)
+
+    async def update_webhook(
+        self, webhook_uuid: NonEmptyString, request: UpdateWebhookRequest
+    ) -> WebhookResponse:
+        """
+        Update webhook configuration.
+
+        Note: Daily.co may not support PATCH for all fields.
+        Common pattern is delete + recreate.
+
+        Reference: https://docs.daily.co/reference/rest-api/webhooks
+
+        Args:
+            webhook_uuid: Webhook UUID to update
+            request: Updated webhook configuration
+
+        Returns:
+            Updated webhook configuration
+
+        Raises:
+            httpx.HTTPStatusError: If API request fails
+        """
+        client = await self._get_client()
+        response = await client.patch(
+            f"{self.base_url}/webhooks/{webhook_uuid}",
+            headers=self.headers,
+            json=request.model_dump(exclude_none=True),
+        )
+
+        data = await self._handle_response(response, "update_webhook")
+        return WebhookResponse(**data)
+
+    async def delete_webhook(self, webhook_uuid: NonEmptyString) -> None:
+        """
+        Delete a webhook.
+
+        Reference: https://docs.daily.co/reference/rest-api/webhooks
+
+        Args:
+            webhook_uuid: Webhook UUID to delete
+
+        Raises:
+            httpx.HTTPStatusError: If webhook not found or deletion fails
+        """
+        client = await self._get_client()
+        response = await client.delete(
+            f"{self.base_url}/webhooks/{webhook_uuid}",
+            headers=self.headers,
+        )
+
+        await self._handle_response(response, "delete_webhook")
+
+    # ============================================================================
+    # HELPER METHODS
+    # ============================================================================
+
+    async def find_webhook_by_url(self, url: NonEmptyString) -> WebhookResponse | None:
+        """
+        Find a webhook by its URL.
+
+        Args:
+            url: Webhook endpoint URL to search for
+
+        Returns:
+            Webhook if found, None otherwise
+        """
+        webhooks = await self.list_webhooks()
+        for webhook in webhooks:
+            if webhook.url == url:
+                return webhook
+        return None
+
+    async def find_webhooks_by_pattern(
+        self, pattern: NonEmptyString
+    ) -> list[WebhookResponse]:
+        """
+        Find webhooks matching a URL pattern (e.g., 'ngrok').
+
+        Args:
+            pattern: String to match in webhook URLs
+
+        Returns:
+            List of matching webhooks
+        """
+        webhooks = await self.list_webhooks()
+        return [wh for wh in webhooks if pattern in wh.url]

server/reflector/dailyco_api/requests.py

@@ -0,0 +1,158 @@
+"""
+Daily.co API Request Models
+
+Reference: https://docs.daily.co/reference/rest-api
+"""
+
+from typing import List, Literal
+
+from pydantic import BaseModel, Field
+
+from reflector.utils.string import NonEmptyString
+
+
+class RecordingsBucketConfig(BaseModel):
+    """
+    S3 bucket configuration for raw-tracks recordings.
+
+    Reference: https://docs.daily.co/reference/rest-api/rooms/create-room
+    """
+
+    bucket_name: NonEmptyString = Field(description="S3 bucket name")
+    bucket_region: NonEmptyString = Field(description="AWS region (e.g., 'us-east-1')")
+    assume_role_arn: NonEmptyString = Field(
+        description="AWS IAM role ARN that Daily.co will assume to write recordings"
+    )
+    allow_api_access: bool = Field(
+        default=True,
+        description="Whether to allow API access to recording metadata",
+    )
+
+
+class RoomProperties(BaseModel):
+    """
+    Room configuration properties.
+    """
+
+    enable_recording: Literal["cloud", "local", "raw-tracks"] | None = Field(
+        default=None,
+        description="Recording mode: 'cloud' for mixed, 'local' for local recording, 'raw-tracks' for multitrack, None to disable",
+    )
+    enable_chat: bool = Field(default=True, description="Enable in-meeting chat")
+    enable_screenshare: bool = Field(default=True, description="Enable screen sharing")
+    start_video_off: bool = Field(
+        default=False, description="Start with video off for all participants"
+    )
+    start_audio_off: bool = Field(
+        default=False, description="Start with audio muted for all participants"
+    )
+    exp: int | None = Field(
+        None, description="Room expiration timestamp (Unix epoch seconds)"
+    )
+    recordings_bucket: RecordingsBucketConfig | None = Field(
+        None, description="S3 bucket configuration for raw-tracks recordings"
+    )
+
+
+class CreateRoomRequest(BaseModel):
+    """
+    Request to create a new Daily.co room.
+
+    Reference: https://docs.daily.co/reference/rest-api/rooms/create-room
+    """
+
+    name: NonEmptyString = Field(description="Room name (must be unique within domain)")
+    privacy: Literal["public", "private"] = Field(
+        default="public", description="Room privacy setting"
+    )
+    properties: RoomProperties = Field(
+        default_factory=RoomProperties, description="Room configuration properties"
+    )
+
+
+class MeetingTokenProperties(BaseModel):
+    """
+    Properties for meeting token creation.
+
+    Reference: https://docs.daily.co/reference/rest-api/meeting-tokens/create-meeting-token
+    """
+
+    room_name: NonEmptyString = Field(description="Room name this token is valid for")
+    user_id: NonEmptyString | None = Field(
+        None, description="User identifier to associate with token"
+    )
+    is_owner: bool = Field(
+        default=False, description="Grant owner privileges to token holder"
+    )
+    start_cloud_recording: bool = Field(
+        default=False, description="Automatically start cloud recording on join"
+    )
+    enable_recording_ui: bool = Field(
+        default=True, description="Show recording controls in UI"
+    )
+    eject_at_token_exp: bool = Field(
+        default=False, description="Eject participant when token expires"
+    )
+    nbf: int | None = Field(
+        None, description="Not-before timestamp (Unix epoch seconds)"
+    )
+    exp: int | None = Field(
+        None, description="Expiration timestamp (Unix epoch seconds)"
+    )
+
+
+class CreateMeetingTokenRequest(BaseModel):
+    """
+    Request to create a meeting token for participant authentication.
+
+    Reference: https://docs.daily.co/reference/rest-api/meeting-tokens/create-meeting-token
+    """
+
+    properties: MeetingTokenProperties = Field(description="Token properties")
+
+
+class CreateWebhookRequest(BaseModel):
+    """
+    Request to create a webhook subscription.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks
+    """
+
+    url: NonEmptyString = Field(description="Webhook endpoint URL (must be HTTPS)")
+    eventTypes: List[
+        Literal[
+            "participant.joined",
+            "participant.left",
+            "recording.started",
+            "recording.ready-to-download",
+            "recording.error",
+        ]
+    ] = Field(
+        description="Array of event types to subscribe to (only events we handle)"
+    )
+    hmac: NonEmptyString = Field(
+        description="Base64-encoded HMAC secret for webhook signature verification"
+    )
+    basicAuth: NonEmptyString | None = Field(
+        None, description="Optional basic auth credentials for webhook endpoint"
+    )
+
+
+class UpdateWebhookRequest(BaseModel):
+    """
+    Request to update an existing webhook.
+
+    Note: Daily.co API may not support PATCH for webhooks.
+    Common pattern is to delete and recreate.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks
+    """
+
+    url: NonEmptyString | None = Field(None, description="New webhook endpoint URL")
+    eventTypes: List[NonEmptyString] | None = Field(
+        None, description="New array of event types"
+    )
+    hmac: NonEmptyString | None = Field(None, description="New HMAC secret")
+    basicAuth: NonEmptyString | None = Field(
+        None, description="New basic auth credentials"
+    )

server/reflector/dailyco_api/responses.py

@@ -0,0 +1,193 @@
+"""
+Daily.co API Response Models
+"""
+
+from typing import Any, Dict, List, Literal
+
+from pydantic import BaseModel, Field
+
+from reflector.dailyco_api.webhooks import DailyTrack
+from reflector.utils.string import NonEmptyString
+
+# not documented in daily; we fill it according to observations
+RecordingStatus = Literal["in-progress", "finished"]
+
+
+class RoomResponse(BaseModel):
+    """
+    Response from room creation or retrieval.
+
+    Reference: https://docs.daily.co/reference/rest-api/rooms/create-room
+    """
+
+    id: NonEmptyString = Field(description="Unique room identifier (UUID)")
+    name: NonEmptyString = Field(description="Room name used in URLs")
+    api_created: bool = Field(description="Whether room was created via API")
+    privacy: Literal["public", "private"] = Field(description="Room privacy setting")
+    url: NonEmptyString = Field(description="Full room URL")
+    created_at: NonEmptyString = Field(description="ISO 8601 creation timestamp")
+    config: Dict[NonEmptyString, Any] = Field(
+        default_factory=dict, description="Room configuration properties"
+    )
+
+
+class RoomPresenceParticipant(BaseModel):
+    """
+    Participant presence information in a room.
+
+    Reference: https://docs.daily.co/reference/rest-api/rooms/get-room-presence
+    """
+
+    room: NonEmptyString = Field(description="Room name")
+    id: NonEmptyString = Field(description="Participant session ID")
+    userId: NonEmptyString | None = Field(None, description="User ID if provided")
+    userName: NonEmptyString | None = Field(None, description="User display name")
+    joinTime: NonEmptyString = Field(description="ISO 8601 join timestamp")
+    duration: int = Field(description="Duration in room (seconds)")
+
+
+class RoomPresenceResponse(BaseModel):
+    """
+    Response from room presence endpoint.
+
+    Reference: https://docs.daily.co/reference/rest-api/rooms/get-room-presence
+    """
+
+    total_count: int = Field(
+        description="Total number of participants currently in room"
+    )
+    data: List[RoomPresenceParticipant] = Field(
+        default_factory=list, description="Array of participant presence data"
+    )
+
+
+class MeetingParticipant(BaseModel):
+    """
+    Historical participant data from a meeting.
+
+    Reference: https://docs.daily.co/reference/rest-api/meetings/get-meeting-participants
+    """
+
+    user_id: NonEmptyString | None = Field(None, description="User identifier")
+    participant_id: NonEmptyString = Field(description="Participant session identifier")
+    user_name: NonEmptyString | None = Field(None, description="User display name")
+    join_time: int = Field(description="Join timestamp (Unix epoch seconds)")
+    duration: int = Field(description="Duration in meeting (seconds)")
+
+
+class MeetingParticipantsResponse(BaseModel):
+    """
+    Response from meeting participants endpoint.
+
+    Reference: https://docs.daily.co/reference/rest-api/meetings/get-meeting-participants
+    """
+
+    data: List[MeetingParticipant] = Field(
+        default_factory=list, description="Array of participant data"
+    )
+
+
+class MeetingResponse(BaseModel):
+    """
+    Response from meeting information endpoint.
+
+    Reference: https://docs.daily.co/reference/rest-api/meetings/get-meeting-information
+    """
+
+    id: NonEmptyString = Field(description="Meeting session identifier (UUID)")
+    room: NonEmptyString = Field(description="Room name where meeting occurred")
+    start_time: int = Field(
+        description="Meeting start Unix timestamp (~15s granularity)"
+    )
+    duration: int = Field(description="Total meeting duration in seconds")
+    ongoing: bool = Field(description="Whether meeting is currently active")
+    max_participants: int = Field(description="Peak concurrent participant count")
+    participants: List[MeetingParticipant] = Field(
+        default_factory=list, description="Array of participant session data"
+    )
+
+
+class RecordingS3Info(BaseModel):
+    """
+    S3 bucket information for a recording.
+
+    Reference: https://docs.daily.co/reference/rest-api/recordings
+    """
+
+    bucket_name: NonEmptyString
+    bucket_region: NonEmptyString
+    endpoint: NonEmptyString | None = None
+
+
+class RecordingResponse(BaseModel):
+    """
+    Response from recording retrieval endpoint.
+
+    Reference: https://docs.daily.co/reference/rest-api/recordings
+    """
+
+    id: NonEmptyString = Field(description="Recording identifier")
+    room_name: NonEmptyString = Field(description="Room where recording occurred")
+    start_ts: int = Field(description="Recording start timestamp (Unix epoch seconds)")
+    status: RecordingStatus = Field(
+        description="Recording status ('in-progress' or 'finished')"
+    )
+    max_participants: int | None = Field(
+        None, description="Maximum participants during recording (may be missing)"
+    )
+    duration: int = Field(description="Recording duration in seconds")
+    share_token: NonEmptyString | None = Field(
+        None, description="Token for sharing recording"
+    )
+    s3: RecordingS3Info | None = Field(None, description="S3 bucket information")
+    tracks: list[DailyTrack] = Field(
+        default_factory=list,
+        description="Track list for raw-tracks recordings (always array, never null)",
+    )
+    # this is not a mistake but a deliberate Daily.co naming decision
+    mtgSessionId: NonEmptyString | None = Field(
+        None, description="Meeting session identifier (may be missing)"
+    )
+
+
+class MeetingTokenResponse(BaseModel):
+    """
+    Response from meeting token creation.
+
+    Reference: https://docs.daily.co/reference/rest-api/meeting-tokens/create-meeting-token
+    """
+
+    token: NonEmptyString = Field(
+        description="JWT meeting token for participant authentication"
+    )
+
+
+class WebhookResponse(BaseModel):
+    """
+    Response from webhook creation or retrieval.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks
+    """
+
+    uuid: NonEmptyString = Field(description="Unique webhook identifier")
+    url: NonEmptyString = Field(description="Webhook endpoint URL")
+    hmac: NonEmptyString | None = Field(
+        None, description="Base64-encoded HMAC secret for signature verification"
+    )
+    basicAuth: NonEmptyString | None = Field(
+        None, description="Basic auth credentials if configured"
+    )
+    eventTypes: List[NonEmptyString] = Field(
+        default_factory=list,
+        description="Array of event types (e.g., ['recording.started', 'participant.joined'])",
+    )
+    state: Literal["ACTIVE", "FAILED"] = Field(
+        description="Webhook state - FAILED after 3+ consecutive failures"
+    )
+    failedCount: int = Field(default=0, description="Number of consecutive failures")
+    lastMomentPushed: NonEmptyString | None = Field(
+        None, description="ISO 8601 timestamp of last successful push"
+    )
+    domainId: NonEmptyString = Field(description="Daily.co domain/account identifier")
+    createdAt: NonEmptyString = Field(description="ISO 8601 creation timestamp")
+    updatedAt: NonEmptyString = Field(description="ISO 8601 last update timestamp")

server/reflector/dailyco_api/webhook_utils.py

@@ -0,0 +1,228 @@
+"""
+Daily.co Webhook Utilities
+
+Utilities for verifying and parsing Daily.co webhook events.
+
+Reference: https://docs.daily.co/reference/rest-api/webhooks
+"""
+
+import base64
+import hmac
+from hashlib import sha256
+
+import structlog
+
+from .webhooks import (
+    DailyWebhookEvent,
+    ParticipantJoinedPayload,
+    ParticipantLeftPayload,
+    RecordingErrorPayload,
+    RecordingReadyToDownloadPayload,
+    RecordingStartedPayload,
+)
+
+logger = structlog.get_logger(__name__)
+
+
+def verify_webhook_signature(
+    body: bytes,
+    signature: str,
+    timestamp: str,
+    webhook_secret: str,
+) -> bool:
+    """
+    Verify Daily.co webhook signature using HMAC-SHA256.
+
+    Daily.co signature verification:
+    1. Base64-decode the webhook secret
+    2. Create signed content: timestamp + '.' + body
+    3. Compute HMAC-SHA256(secret, signed_content)
+    4. Base64-encode the result
+    5. Compare with provided signature using constant-time comparison
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks
+
+    Args:
+        body: Raw request body bytes
+        signature: X-Webhook-Signature header value
+        timestamp: X-Webhook-Timestamp header value
+        webhook_secret: Base64-encoded HMAC secret
+
+    Returns:
+        True if signature is valid, False otherwise
+
+    Example:
+        >>> body = b'{"version":"1.0.0","type":"participant.joined",...}'
+        >>> signature = "abc123..."
+        >>> timestamp = "1234567890"
+        >>> secret = "your-base64-secret"
+        >>> is_valid = verify_webhook_signature(body, signature, timestamp, secret)
+    """
+    if not signature or not timestamp or not webhook_secret:
+        logger.warning(
+            "Missing required data for webhook verification",
+            has_signature=bool(signature),
+            has_timestamp=bool(timestamp),
+            has_secret=bool(webhook_secret),
+        )
+        return False
+
+    try:
+        secret_bytes = base64.b64decode(webhook_secret)
+        signed_content = timestamp.encode() + b"." + body
+        expected = hmac.new(secret_bytes, signed_content, sha256).digest()
+        expected_b64 = base64.b64encode(expected).decode()
+
+        # Constant-time comparison to prevent timing attacks
+        return hmac.compare_digest(expected_b64, signature)
+
+    except (base64.binascii.Error, ValueError, TypeError, UnicodeDecodeError) as e:
+        logger.error(
+            "Webhook signature verification failed",
+            error=str(e),
+            error_type=type(e).__name__,
+        )
+        return False
+
+
+def extract_room_name(event: DailyWebhookEvent) -> str | None:
+    """
+    Extract room name from Daily.co webhook event payload.
+
+    Args:
+        event: Parsed webhook event
+
+    Returns:
+        Room name if present and is a string, None otherwise
+
+    Example:
+        >>> event = DailyWebhookEvent(**webhook_payload)
+        >>> room_name = extract_room_name(event)
+    """
+    room = event.payload.get("room_name")
+    # Ensure we return a string, not any falsy value that might be in payload
+    return room if isinstance(room, str) else None
+
+
+def parse_participant_joined(event: DailyWebhookEvent) -> ParticipantJoinedPayload:
+    """
+    Parse participant.joined webhook event payload.
+
+    Args:
+        event: Webhook event with type "participant.joined"
+
+    Returns:
+        Parsed participant joined payload
+
+    Raises:
+        pydantic.ValidationError: If payload doesn't match expected schema
+    """
+    return ParticipantJoinedPayload(**event.payload)
+
+
+def parse_participant_left(event: DailyWebhookEvent) -> ParticipantLeftPayload:
+    """
+    Parse participant.left webhook event payload.
+
+    Args:
+        event: Webhook event with type "participant.left"
+
+    Returns:
+        Parsed participant left payload
+
+    Raises:
+        pydantic.ValidationError: If payload doesn't match expected schema
+    """
+    return ParticipantLeftPayload(**event.payload)
+
+
+def parse_recording_started(event: DailyWebhookEvent) -> RecordingStartedPayload:
+    """
+    Parse recording.started webhook event payload.
+
+    Args:
+        event: Webhook event with type "recording.started"
+
+    Returns:
+        Parsed recording started payload
+
+    Raises:
+        pydantic.ValidationError: If payload doesn't match expected schema
+    """
+    return RecordingStartedPayload(**event.payload)
+
+
+def parse_recording_ready(
+    event: DailyWebhookEvent,
+) -> RecordingReadyToDownloadPayload:
+    """
+    Parse recording.ready-to-download webhook event payload.
+
+    This event is sent when raw-tracks recordings are complete and uploaded to S3.
+    The payload includes a 'tracks' array with individual audio/video files.
+
+    Args:
+        event: Webhook event with type "recording.ready-to-download"
+
+    Returns:
+        Parsed recording ready payload with tracks array
+
+    Raises:
+        pydantic.ValidationError: If payload doesn't match expected schema
+
+    Example:
+        >>> event = DailyWebhookEvent(**webhook_payload)
+        >>> if event.type == "recording.ready-to-download":
+        ...     payload = parse_recording_ready(event)
+        ...     audio_tracks = [t for t in payload.tracks if t.type == "audio"]
+    """
+    return RecordingReadyToDownloadPayload(**event.payload)
+
+
+def parse_recording_error(event: DailyWebhookEvent) -> RecordingErrorPayload:
+    """
+    Parse recording.error webhook event payload.
+
+    Args:
+        event: Webhook event with type "recording.error"
+
+    Returns:
+        Parsed recording error payload
+
+    Raises:
+        pydantic.ValidationError: If payload doesn't match expected schema
+    """
+    return RecordingErrorPayload(**event.payload)
+
+
+WEBHOOK_PARSERS = {
+    "participant.joined": parse_participant_joined,
+    "participant.left": parse_participant_left,
+    "recording.started": parse_recording_started,
+    "recording.ready-to-download": parse_recording_ready,
+    "recording.error": parse_recording_error,
+}
+
+
+def parse_webhook_payload(event: DailyWebhookEvent):
+    """
+    Parse webhook event payload based on event type.
+
+    Args:
+        event: Webhook event
+
+    Returns:
+        Typed payload model based on event type, or raw dict if unknown
+
+    Example:
+        >>> event = DailyWebhookEvent(**webhook_payload)
+        >>> payload = parse_webhook_payload(event)
+        >>> if isinstance(payload, ParticipantJoinedPayload):
+        ...     print(f"User {payload.user_name} joined")
+    """
+    parser = WEBHOOK_PARSERS.get(event.type)
+    if parser:
+        return parser(event)
+    else:
+        logger.warning("Unknown webhook event type", event_type=event.type)
+        return event.payload

server/reflector/dailyco_api/webhooks.py

@@ -0,0 +1,271 @@
+"""
+Daily.co Webhook Event Models
+
+Reference: https://docs.daily.co/reference/rest-api/webhooks
+"""
+
+from typing import Annotated, Any, Dict, Literal, Union
+
+from pydantic import BaseModel, Field, field_validator
+
+from reflector.utils.string import NonEmptyString
+
+
+def normalize_timestamp_to_int(v):
+    """
+    Normalize float timestamps to int by truncating decimal part.
+
+    Daily.co sometimes sends timestamps as floats (e.g., 1708972279.96).
+    Pydantic expects int for fields typed as `int`.
+    """
+    if v is None:
+        return v
+    if isinstance(v, float):
+        return int(v)
+    return v
+
+
+WebhookEventType = Literal[
+    "participant.joined",
+    "participant.left",
+    "recording.started",
+    "recording.ready-to-download",
+    "recording.error",
+]
+
+
+class DailyTrack(BaseModel):
+    """
+    Individual audio or video track from a multitrack recording.
+
+    Reference: https://docs.daily.co/reference/rest-api/recordings
+    """
+
+    type: Literal["audio", "video"]
+    s3Key: NonEmptyString = Field(description="S3 object key for the track file")
+    size: int = Field(description="File size in bytes")
+
+
+class DailyWebhookEvent(BaseModel):
+    """
+    Base structure for all Daily.co webhook events.
+    All events share five common fields documented below.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks
+    """
+
+    version: NonEmptyString = Field(
+        description="Represents the version of the event. This uses semantic versioning to inform a consumer if the payload has introduced any breaking changes"
+    )
+    type: WebhookEventType = Field(
+        description="Represents the type of the event described in the payload"
+    )
+    id: NonEmptyString = Field(
+        description="An identifier representing this specific event"
+    )
+    payload: Dict[NonEmptyString, Any] = Field(
+        description="An object representing the event, whose fields are described in the corresponding payload class"
+    )
+    event_ts: int = Field(
+        description="Documenting when the webhook itself was sent. This timestamp is different than the time of the event the webhook describes. For example, a recording.started event will contain a start_ts timestamp of when the actual recording started, and a slightly later event_ts timestamp indicating when the webhook event was sent"
+    )
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class ParticipantJoinedPayload(BaseModel):
+    """
+    Payload for participant.joined webhook event.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks/events/participant-joined
+    """
+
+    room_name: NonEmptyString | None = Field(None, description="Daily.co room name")
+    session_id: NonEmptyString = Field(description="Daily.co session identifier")
+    user_id: NonEmptyString = Field(description="User identifier (may be encoded)")
+    user_name: NonEmptyString | None = Field(None, description="User display name")
+    joined_at: int = Field(description="Join timestamp in Unix epoch seconds")
+
+    _normalize_joined_at = field_validator("joined_at", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class ParticipantLeftPayload(BaseModel):
+    """
+    Payload for participant.left webhook event.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks/events/participant-left
+    """
+
+    room_name: NonEmptyString | None = Field(None, description="Daily.co room name")
+    session_id: NonEmptyString = Field(description="Daily.co session identifier")
+    user_id: NonEmptyString = Field(description="User identifier (may be encoded)")
+    user_name: NonEmptyString | None = Field(None, description="User display name")
+    joined_at: int = Field(description="Join timestamp in Unix epoch seconds")
+    duration: int | None = Field(
+        None, description="Duration of participation in seconds"
+    )
+
+    _normalize_joined_at = field_validator("joined_at", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingStartedPayload(BaseModel):
+    """
+    Payload for recording.started webhook event.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks/events/recording-started
+    """
+
+    room_name: NonEmptyString | None = Field(None, description="Daily.co room name")
+    recording_id: NonEmptyString = Field(description="Recording identifier")
+    start_ts: int | None = Field(None, description="Recording start timestamp")
+
+    _normalize_start_ts = field_validator("start_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingReadyToDownloadPayload(BaseModel):
+    """
+    Payload for recording.ready-to-download webhook event.
+    This is sent when raw-tracks recordings are complete and uploaded to S3.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks/events/recording-ready-to-download
+    """
+
+    type: Literal["cloud", "raw-tracks"] = Field(
+        description="The type of recording that was generated"
+    )
+    recording_id: NonEmptyString = Field(
+        description="An ID identifying the recording that was generated"
+    )
+    room_name: NonEmptyString = Field(
+        description="The name of the room where the recording was made"
+    )
+    start_ts: int = Field(
+        description="The Unix epoch time in seconds representing when the recording started"
+    )
+    status: Literal["finished"] = Field(
+        description="The status of the given recording (always 'finished' in ready-to-download webhook, see RecordingStatus in responses.py for full API statuses)"
+    )
+    max_participants: int = Field(
+        description="The number of participants on the call that were recorded"
+    )
+    duration: int = Field(description="The duration in seconds of the call")
+    s3_key: NonEmptyString = Field(
+        description="The location of the recording in the provided S3 bucket"
+    )
+    share_token: NonEmptyString | None = Field(
+        None, description="undocumented documented secret field"
+    )
+    tracks: list[DailyTrack] | None = Field(
+        None,
+        description="If the recording is a raw-tracks recording, a tracks field will be provided. If role permissions have been removed, the tracks field may be null",
+    )
+
+    _normalize_start_ts = field_validator("start_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingErrorPayload(BaseModel):
+    """
+    Payload for recording.error webhook event.
+
+    Reference: https://docs.daily.co/reference/rest-api/webhooks/events/recording-error
+    """
+
+    action: Literal["clourd-recording-err", "cloud-recording-error"] = Field(
+        description="A string describing the event that was emitted (both variants are documented)"
+    )
+    error_msg: NonEmptyString = Field(description="The error message returned")
+    instance_id: NonEmptyString = Field(
+        description="The recording instance ID that was passed into the start recording command"
+    )
+    room_name: NonEmptyString = Field(
+        description="The name of the room where the recording was made"
+    )
+    timestamp: int = Field(
+        description="The Unix epoch time in seconds representing when the error was emitted"
+    )
+
+    _normalize_timestamp = field_validator("timestamp", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class ParticipantJoinedEvent(BaseModel):
+    version: NonEmptyString
+    type: Literal["participant.joined"]
+    id: NonEmptyString
+    payload: ParticipantJoinedPayload
+    event_ts: int
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class ParticipantLeftEvent(BaseModel):
+    version: NonEmptyString
+    type: Literal["participant.left"]
+    id: NonEmptyString
+    payload: ParticipantLeftPayload
+    event_ts: int
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingStartedEvent(BaseModel):
+    version: NonEmptyString
+    type: Literal["recording.started"]
+    id: NonEmptyString
+    payload: RecordingStartedPayload
+    event_ts: int
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingReadyEvent(BaseModel):
+    version: NonEmptyString
+    type: Literal["recording.ready-to-download"]
+    id: NonEmptyString
+    payload: RecordingReadyToDownloadPayload
+    event_ts: int
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+class RecordingErrorEvent(BaseModel):
+    version: NonEmptyString
+    type: Literal["recording.error"]
+    id: NonEmptyString
+    payload: RecordingErrorPayload
+    event_ts: int
+
+    _normalize_event_ts = field_validator("event_ts", mode="before")(
+        normalize_timestamp_to_int
+    )
+
+
+DailyWebhookEventUnion = Annotated[
+    Union[
+        ParticipantJoinedEvent,
+        ParticipantLeftEvent,
+        RecordingStartedEvent,
+        RecordingReadyEvent,
+        RecordingErrorEvent,
+    ],
+    Field(discriminator="type"),
+]

server/reflector/db/__init__.py

@@ -24,10 +24,14 @@ def get_database() -> databases.Database:
 
 
 # import models
+import reflector.db.calendar_events  # noqa
+import reflector.db.daily_participant_sessions  # noqa
 import reflector.db.meetings  # noqa
 import reflector.db.recordings  # noqa
 import reflector.db.rooms  # noqa
 import reflector.db.transcripts  # noqa
+import reflector.db.user_api_keys  # noqa
+import reflector.db.users  # noqa
 
 kwargs = {}
 if "postgres" not in settings.DATABASE_URL:

server/reflector/db/calendar_events.py

@@ -0,0 +1,187 @@
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+import sqlalchemy as sa
+from pydantic import BaseModel, Field
+from sqlalchemy.dialects.postgresql import JSONB
+
+from reflector.db import get_database, metadata
+from reflector.utils import generate_uuid4
+
+calendar_events = sa.Table(
+    "calendar_event",
+    metadata,
+    sa.Column("id", sa.String, primary_key=True),
+    sa.Column(
+        "room_id",
+        sa.String,
+        sa.ForeignKey("room.id", ondelete="CASCADE", name="fk_calendar_event_room_id"),
+        nullable=False,
+    ),
+    sa.Column("ics_uid", sa.Text, nullable=False),
+    sa.Column("title", sa.Text),
+    sa.Column("description", sa.Text),
+    sa.Column("start_time", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("end_time", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("attendees", JSONB),
+    sa.Column("location", sa.Text),
+    sa.Column("ics_raw_data", sa.Text),
+    sa.Column("last_synced", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("is_deleted", sa.Boolean, nullable=False, server_default=sa.false()),
+    sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+    sa.UniqueConstraint("room_id", "ics_uid", name="uq_room_calendar_event"),
+    sa.Index("idx_calendar_event_room_start", "room_id", "start_time"),
+    sa.Index(
+        "idx_calendar_event_deleted",
+        "is_deleted",
+        postgresql_where=sa.text("NOT is_deleted"),
+    ),
+)
+
+
+class CalendarEvent(BaseModel):
+    id: str = Field(default_factory=generate_uuid4)
+    room_id: str
+    ics_uid: str
+    title: str | None = None
+    description: str | None = None
+    start_time: datetime
+    end_time: datetime
+    attendees: list[dict[str, Any]] | None = None
+    location: str | None = None
+    ics_raw_data: str | None = None
+    last_synced: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    is_deleted: bool = False
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class CalendarEventController:
+    async def get_by_room(
+        self,
+        room_id: str,
+        include_deleted: bool = False,
+        start_after: datetime | None = None,
+        end_before: datetime | None = None,
+    ) -> list[CalendarEvent]:
+        query = calendar_events.select().where(calendar_events.c.room_id == room_id)
+
+        if not include_deleted:
+            query = query.where(calendar_events.c.is_deleted == False)
+
+        if start_after:
+            query = query.where(calendar_events.c.start_time >= start_after)
+
+        if end_before:
+            query = query.where(calendar_events.c.end_time <= end_before)
+
+        query = query.order_by(calendar_events.c.start_time.asc())
+
+        results = await get_database().fetch_all(query)
+        return [CalendarEvent(**result) for result in results]
+
+    async def get_upcoming(
+        self, room_id: str, minutes_ahead: int = 120
+    ) -> list[CalendarEvent]:
+        """Get upcoming events for a room within the specified minutes, including currently happening events."""
+        now = datetime.now(timezone.utc)
+        future_time = now + timedelta(minutes=minutes_ahead)
+
+        query = (
+            calendar_events.select()
+            .where(
+                sa.and_(
+                    calendar_events.c.room_id == room_id,
+                    calendar_events.c.is_deleted == False,
+                    calendar_events.c.start_time <= future_time,
+                    calendar_events.c.end_time >= now,
+                )
+            )
+            .order_by(calendar_events.c.start_time.asc())
+        )
+
+        results = await get_database().fetch_all(query)
+        return [CalendarEvent(**result) for result in results]
+
+    async def get_by_id(self, event_id: str) -> CalendarEvent | None:
+        query = calendar_events.select().where(calendar_events.c.id == event_id)
+        result = await get_database().fetch_one(query)
+        return CalendarEvent(**result) if result else None
+
+    async def get_by_ics_uid(self, room_id: str, ics_uid: str) -> CalendarEvent | None:
+        query = calendar_events.select().where(
+            sa.and_(
+                calendar_events.c.room_id == room_id,
+                calendar_events.c.ics_uid == ics_uid,
+            )
+        )
+        result = await get_database().fetch_one(query)
+        return CalendarEvent(**result) if result else None
+
+    async def upsert(self, event: CalendarEvent) -> CalendarEvent:
+        existing = await self.get_by_ics_uid(event.room_id, event.ics_uid)
+
+        if existing:
+            event.id = existing.id
+            event.created_at = existing.created_at
+            event.updated_at = datetime.now(timezone.utc)
+
+            query = (
+                calendar_events.update()
+                .where(calendar_events.c.id == existing.id)
+                .values(**event.model_dump())
+            )
+        else:
+            query = calendar_events.insert().values(**event.model_dump())
+
+        await get_database().execute(query)
+        return event
+
+    async def soft_delete_missing(
+        self, room_id: str, current_ics_uids: list[str]
+    ) -> int:
+        """Soft delete future events that are no longer in the calendar."""
+        now = datetime.now(timezone.utc)
+
+        select_query = calendar_events.select().where(
+            sa.and_(
+                calendar_events.c.room_id == room_id,
+                calendar_events.c.start_time > now,
+                calendar_events.c.is_deleted == False,
+                calendar_events.c.ics_uid.notin_(current_ics_uids)
+                if current_ics_uids
+                else True,
+            )
+        )
+
+        to_delete = await get_database().fetch_all(select_query)
+        delete_count = len(to_delete)
+
+        if delete_count > 0:
+            update_query = (
+                calendar_events.update()
+                .where(
+                    sa.and_(
+                        calendar_events.c.room_id == room_id,
+                        calendar_events.c.start_time > now,
+                        calendar_events.c.is_deleted == False,
+                        calendar_events.c.ics_uid.notin_(current_ics_uids)
+                        if current_ics_uids
+                        else True,
+                    )
+                )
+                .values(is_deleted=True, updated_at=now)
+            )
+
+            await get_database().execute(update_query)
+
+        return delete_count
+
+    async def delete_by_room(self, room_id: str) -> int:
+        query = calendar_events.delete().where(calendar_events.c.room_id == room_id)
+        result = await get_database().execute(query)
+        return result.rowcount
+
+
+calendar_events_controller = CalendarEventController()

server/reflector/db/daily_participant_sessions.py

@@ -0,0 +1,229 @@
+"""Daily.co participant session tracking.
+
+Stores webhook data for participant.joined and participant.left events to provide
+historical session information (Daily.co API only returns current participants).
+"""
+
+from datetime import datetime
+
+import sqlalchemy as sa
+from pydantic import BaseModel
+from sqlalchemy.dialects.postgresql import insert
+
+from reflector.db import get_database, metadata
+from reflector.utils.string import NonEmptyString
+
+daily_participant_sessions = sa.Table(
+    "daily_participant_session",
+    metadata,
+    sa.Column("id", sa.String, primary_key=True),
+    sa.Column(
+        "meeting_id",
+        sa.String,
+        sa.ForeignKey("meeting.id", ondelete="CASCADE"),
+        nullable=False,
+    ),
+    sa.Column(
+        "room_id",
+        sa.String,
+        sa.ForeignKey("room.id", ondelete="CASCADE"),
+        nullable=False,
+    ),
+    sa.Column("session_id", sa.String, nullable=False),
+    sa.Column("user_id", sa.String, nullable=True),
+    sa.Column("user_name", sa.String, nullable=False),
+    sa.Column("joined_at", sa.DateTime(timezone=True), nullable=False),
+    sa.Column("left_at", sa.DateTime(timezone=True), nullable=True),
+    sa.Index("idx_daily_session_meeting_left", "meeting_id", "left_at"),
+    sa.Index("idx_daily_session_room", "room_id"),
+)
+
+
+class DailyParticipantSession(BaseModel):
+    """Daily.co participant session record.
+
+    Tracks when a participant joined and left a meeting. Populated from webhooks:
+    - participant.joined: Creates record with left_at=None
+    - participant.left: Updates record with left_at
+
+    ID format: {meeting_id}:{user_id}:{joined_at_ms}
+    - Ensures idempotency (duplicate webhooks don't create duplicates)
+    - Allows same user to rejoin (different joined_at = different session)
+
+    Duration is calculated as: left_at - joined_at (not stored)
+    """
+
+    id: NonEmptyString
+    meeting_id: NonEmptyString
+    room_id: NonEmptyString
+    session_id: NonEmptyString  # Daily.co's session_id (identifies room session)
+    user_id: NonEmptyString | None = None
+    user_name: str
+    joined_at: datetime
+    left_at: datetime | None = None
+
+
+class DailyParticipantSessionController:
+    """Controller for Daily.co participant session persistence."""
+
+    async def get_by_id(self, id: str) -> DailyParticipantSession | None:
+        """Get a session by its ID."""
+        query = daily_participant_sessions.select().where(
+            daily_participant_sessions.c.id == id
+        )
+        result = await get_database().fetch_one(query)
+        return DailyParticipantSession(**result) if result else None
+
+    async def get_open_session(
+        self, meeting_id: NonEmptyString, session_id: NonEmptyString
+    ) -> DailyParticipantSession | None:
+        """Get the open (not left) session for a user in a meeting."""
+        query = daily_participant_sessions.select().where(
+            sa.and_(
+                daily_participant_sessions.c.meeting_id == meeting_id,
+                daily_participant_sessions.c.session_id == session_id,
+                daily_participant_sessions.c.left_at.is_(None),
+            )
+        )
+        results = await get_database().fetch_all(query)
+
+        if len(results) > 1:
+            raise ValueError(
+                f"Multiple open sessions for daily session {session_id} in meeting {meeting_id}: "
+                f"found {len(results)} sessions"
+            )
+
+        return DailyParticipantSession(**results[0]) if results else None
+
+    async def upsert_joined(self, session: DailyParticipantSession) -> None:
+        """Insert or update when participant.joined webhook arrives.
+
+        Idempotent: Duplicate webhooks with same ID are safely ignored.
+        Out-of-order: If left webhook arrived first, preserves left_at.
+        """
+        query = insert(daily_participant_sessions).values(**session.model_dump())
+        query = query.on_conflict_do_update(
+            index_elements=["id"],
+            set_={"user_name": session.user_name},
+        )
+        await get_database().execute(query)
+
+    async def upsert_left(self, session: DailyParticipantSession) -> None:
+        """Update session when participant.left webhook arrives.
+
+        Finds the open session for this user in this meeting and updates left_at.
+        Works around Daily.co webhook timestamp inconsistency (joined_at differs by ~4ms between webhooks).
+
+        Handles three cases:
+        1. Normal flow: open session exists → updates left_at
+        2. Out-of-order: left arrives first → creates new record with left data
+        3. Duplicate: left arrives again → idempotent (DB trigger prevents left_at modification)
+        """
+        if session.left_at is None:
+            raise ValueError("left_at is required for upsert_left")
+
+        if session.left_at <= session.joined_at:
+            raise ValueError(
+                f"left_at ({session.left_at}) must be after joined_at ({session.joined_at})"
+            )
+
+        # Find existing open session (works around timestamp mismatch in webhooks)
+        existing = await self.get_open_session(session.meeting_id, session.session_id)
+
+        if existing:
+            # Update existing open session
+            query = (
+                daily_participant_sessions.update()
+                .where(daily_participant_sessions.c.id == existing.id)
+                .values(left_at=session.left_at)
+            )
+            await get_database().execute(query)
+        else:
+            # Out-of-order or first webhook: insert new record
+            query = insert(daily_participant_sessions).values(**session.model_dump())
+            query = query.on_conflict_do_nothing(index_elements=["id"])
+            await get_database().execute(query)
+
+    async def get_by_meeting(self, meeting_id: str) -> list[DailyParticipantSession]:
+        """Get all participant sessions for a meeting (active and ended)."""
+        query = daily_participant_sessions.select().where(
+            daily_participant_sessions.c.meeting_id == meeting_id
+        )
+        results = await get_database().fetch_all(query)
+        return [DailyParticipantSession(**result) for result in results]
+
+    async def get_active_by_meeting(
+        self, meeting_id: str
+    ) -> list[DailyParticipantSession]:
+        """Get only active (not left) participant sessions for a meeting."""
+        query = daily_participant_sessions.select().where(
+            sa.and_(
+                daily_participant_sessions.c.meeting_id == meeting_id,
+                daily_participant_sessions.c.left_at.is_(None),
+            )
+        )
+        results = await get_database().fetch_all(query)
+        return [DailyParticipantSession(**result) for result in results]
+
+    async def get_all_sessions_for_meeting(
+        self, meeting_id: NonEmptyString
+    ) -> dict[NonEmptyString, DailyParticipantSession]:
+        query = daily_participant_sessions.select().where(
+            daily_participant_sessions.c.meeting_id == meeting_id
+        )
+        results = await get_database().fetch_all(query)
+        # TODO DailySessionId custom type
+        return {row["session_id"]: DailyParticipantSession(**row) for row in results}
+
+    async def batch_upsert_sessions(
+        self, sessions: list[DailyParticipantSession]
+    ) -> None:
+        """Upsert multiple sessions in single query.
+
+        Uses ON CONFLICT for idempotency. Updates user_name on conflict since they may change it during a meeting.
+
+        """
+        if not sessions:
+            return
+
+        values = [session.model_dump() for session in sessions]
+        query = insert(daily_participant_sessions).values(values)
+        query = query.on_conflict_do_update(
+            index_elements=["id"],
+            set_={
+                # Preserve existing left_at to prevent race conditions
+                "left_at": sa.func.coalesce(
+                    daily_participant_sessions.c.left_at,
+                    query.excluded.left_at,
+                ),
+                "user_name": query.excluded.user_name,
+            },
+        )
+        await get_database().execute(query)
+
+    async def batch_close_sessions(
+        self, session_ids: list[NonEmptyString], left_at: datetime
+    ) -> None:
+        """Mark multiple sessions as left in single query.
+
+        Only updates sessions where left_at is NULL (protects already-closed sessions).
+
+        Left_at mismatch for existing sessions is ignored, assumed to be not important issue if ever happens.
+        """
+        if not session_ids:
+            return
+
+        query = (
+            daily_participant_sessions.update()
+            .where(
+                sa.and_(
+                    daily_participant_sessions.c.id.in_(session_ids),
+                    daily_participant_sessions.c.left_at.is_(None),
+                )
+            )
+            .values(left_at=left_at)
+        )
+        await get_database().execute(query)
+
+
+daily_participant_sessions_controller = DailyParticipantSessionController()

server/reflector/db/meetings.py

@@ -1,13 +1,15 @@
 from datetime import datetime
-from typing import Literal
+from typing import Any, Literal
 
 import sqlalchemy as sa
-from fastapi import HTTPException
 from pydantic import BaseModel, Field
+from sqlalchemy.dialects.postgresql import JSONB
 
 from reflector.db import get_database, metadata
 from reflector.db.rooms import Room
+from reflector.schemas.platform import WHEREBY_PLATFORM, Platform
 from reflector.utils import generate_uuid4
+from reflector.utils.string import assert_equal
 
 meetings = sa.Table(
     "meeting",
@@ -18,8 +20,12 @@ meetings = sa.Table(
     sa.Column("host_room_url", sa.String),
     sa.Column("start_date", sa.DateTime(timezone=True)),
     sa.Column("end_date", sa.DateTime(timezone=True)),
-    sa.Column("user_id", sa.String),
-    sa.Column("room_id", sa.String),
+    sa.Column(
+        "room_id",
+        sa.String,
+        sa.ForeignKey("room.id", ondelete="CASCADE"),
+        nullable=True,
+    ),
     sa.Column("is_locked", sa.Boolean, nullable=False, server_default=sa.false()),
     sa.Column("room_mode", sa.String, nullable=False, server_default="normal"),
     sa.Column("recording_type", sa.String, nullable=False, server_default="cloud"),
@@ -41,20 +47,36 @@ meetings = sa.Table(
         nullable=False,
         server_default=sa.true(),
     ),
-    sa.Index("idx_meeting_room_id", "room_id"),
-    sa.Index(
-        "idx_one_active_meeting_per_room",
-        "room_id",
-        unique=True,
-        postgresql_where=sa.text("is_active = true"),
+    sa.Column(
+        "calendar_event_id",
+        sa.String,
+        sa.ForeignKey(
+            "calendar_event.id",
+            ondelete="SET NULL",
+            name="fk_meeting_calendar_event_id",
+        ),
     ),
+    sa.Column("calendar_metadata", JSONB),
+    sa.Column(
+        "platform",
+        sa.String,
+        nullable=False,
+        server_default=assert_equal(WHEREBY_PLATFORM, "whereby"),
+    ),
+    sa.Index("idx_meeting_room_id", "room_id"),
+    sa.Index("idx_meeting_calendar_event", "calendar_event_id"),
 )
 
 meeting_consent = sa.Table(
     "meeting_consent",
     metadata,
     sa.Column("id", sa.String, primary_key=True),
-    sa.Column("meeting_id", sa.String, sa.ForeignKey("meeting.id"), nullable=False),
+    sa.Column(
+        "meeting_id",
+        sa.String,
+        sa.ForeignKey("meeting.id", ondelete="CASCADE"),
+        nullable=False,
+    ),
     sa.Column("user_id", sa.String),
     sa.Column("consent_given", sa.Boolean, nullable=False),
     sa.Column("consent_timestamp", sa.DateTime(timezone=True), nullable=False),
@@ -76,15 +98,18 @@ class Meeting(BaseModel):
     host_room_url: str
     start_date: datetime
     end_date: datetime
-    user_id: str | None = None
-    room_id: str | None = None
+    room_id: str | None
     is_locked: bool = False
     room_mode: Literal["normal", "group"] = "normal"
     recording_type: Literal["none", "local", "cloud"] = "cloud"
-    recording_trigger: Literal[
+    recording_trigger: Literal[  # whereby-specific
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     num_clients: int = 0
+    is_active: bool = True
+    calendar_event_id: str | None = None
+    calendar_metadata: dict[str, Any] | None = None
+    platform: Platform = WHEREBY_PLATFORM
 
 
 class MeetingController:
@@ -96,12 +121,10 @@ class MeetingController:
         host_room_url: str,
         start_date: datetime,
         end_date: datetime,
-        user_id: str,
         room: Room,
+        calendar_event_id: str | None = None,
+        calendar_metadata: dict[str, Any] | None = None,
     ):
-        """
-        Create a new meeting
-        """
         meeting = Meeting(
             id=id,
             room_name=room_name,
@@ -109,41 +132,49 @@ class MeetingController:
             host_room_url=host_room_url,
             start_date=start_date,
             end_date=end_date,
-            user_id=user_id,
             room_id=room.id,
             is_locked=room.is_locked,
             room_mode=room.room_mode,
             recording_type=room.recording_type,
             recording_trigger=room.recording_trigger,
+            calendar_event_id=calendar_event_id,
+            calendar_metadata=calendar_metadata,
+            platform=room.platform,
         )
         query = meetings.insert().values(**meeting.model_dump())
         await get_database().execute(query)
         return meeting
 
-    async def get_all_active(self) -> list[Meeting]:
-        """
-        Get active meetings.
-        """
-        query = meetings.select().where(meetings.c.is_active)
-        return await get_database().fetch_all(query)
+    async def get_all_active(self, platform: str | None = None) -> list[Meeting]:
+        conditions = [meetings.c.is_active]
+        if platform is not None:
+            conditions.append(meetings.c.platform == platform)
+        query = meetings.select().where(sa.and_(*conditions))
+        results = await get_database().fetch_all(query)
+        return [Meeting(**result) for result in results]
 
     async def get_by_room_name(
         self,
         room_name: str,
-    ) -> Meeting:
+    ) -> Meeting | None:
         """
         Get a meeting by room name.
+        For backward compatibility, returns the most recent meeting.
         """
-        query = meetings.select().where(meetings.c.room_name == room_name)
+        query = (
+            meetings.select()
+            .where(meetings.c.room_name == room_name)
+            .order_by(meetings.c.end_date.desc())
+        )
         result = await get_database().fetch_one(query)
         if not result:
             return None
-
         return Meeting(**result)
 
-    async def get_active(self, room: Room, current_time: datetime) -> Meeting:
+    async def get_active(self, room: Room, current_time: datetime) -> Meeting | None:
         """
         Get latest active meeting for a room.
+        For backward compatibility, returns the most recent active meeting.
         """
         end_date = getattr(meetings.c, "end_date")
         query = (
@@ -160,40 +191,97 @@ class MeetingController:
         result = await get_database().fetch_one(query)
         if not result:
             return None
-
         return Meeting(**result)
 
-    async def get_by_id(self, meeting_id: str, **kwargs) -> Meeting | None:
+    async def get_all_active_for_room(
+        self, room: Room, current_time: datetime
+    ) -> list[Meeting]:
+        end_date = getattr(meetings.c, "end_date")
+        query = (
+            meetings.select()
+            .where(
+                sa.and_(
+                    meetings.c.room_id == room.id,
+                    meetings.c.end_date > current_time,
+                    meetings.c.is_active,
+                )
+            )
+            .order_by(end_date.desc())
+        )
+        results = await get_database().fetch_all(query)
+        return [Meeting(**result) for result in results]
+
+    async def get_active_by_calendar_event(
+        self, room: Room, calendar_event_id: str, current_time: datetime
+    ) -> Meeting | None:
         """
-        Get a meeting by id
+        Get active meeting for a specific calendar event.
         """
-        query = meetings.select().where(meetings.c.id == meeting_id)
+        query = meetings.select().where(
+            sa.and_(
+                meetings.c.room_id == room.id,
+                meetings.c.calendar_event_id == calendar_event_id,
+                meetings.c.end_date > current_time,
+                meetings.c.is_active,
+            )
+        )
         result = await get_database().fetch_one(query)
         if not result:
             return None
         return Meeting(**result)
 
-    async def get_by_id_for_http(self, meeting_id: str, user_id: str | None) -> Meeting:
-        """
-        Get a meeting by ID for HTTP request.
-
-        If not found, it will raise a 404 error.
-        """
+    async def get_by_id(
+        self, meeting_id: str, room: Room | None = None
+    ) -> Meeting | None:
         query = meetings.select().where(meetings.c.id == meeting_id)
+
+        if room:
+            query = query.where(meetings.c.room_id == room.id)
+
         result = await get_database().fetch_one(query)
         if not result:
-            raise HTTPException(status_code=404, detail="Meeting not found")
+            return None
+        return Meeting(**result)
 
-        meeting = Meeting(**result)
-        if result["user_id"] != user_id:
-            meeting.host_room_url = ""
-
-        return meeting
+    async def get_by_calendar_event(
+        self, calendar_event_id: str, room: Room
+    ) -> Meeting | None:
+        query = meetings.select().where(
+            meetings.c.calendar_event_id == calendar_event_id
+        )
+        if room:
+            query = query.where(meetings.c.room_id == room.id)
+        result = await get_database().fetch_one(query)
+        if not result:
+            return None
+        return Meeting(**result)
 
     async def update_meeting(self, meeting_id: str, **kwargs):
         query = meetings.update().where(meetings.c.id == meeting_id).values(**kwargs)
         await get_database().execute(query)
 
+    async def increment_num_clients(self, meeting_id: str) -> None:
+        """Atomically increment participant count."""
+        query = (
+            meetings.update()
+            .where(meetings.c.id == meeting_id)
+            .values(num_clients=meetings.c.num_clients + 1)
+        )
+        await get_database().execute(query)
+
+    async def decrement_num_clients(self, meeting_id: str) -> None:
+        """Atomically decrement participant count (min 0)."""
+        query = (
+            meetings.update()
+            .where(meetings.c.id == meeting_id)
+            .values(
+                num_clients=sa.case(
+                    (meetings.c.num_clients > 0, meetings.c.num_clients - 1), else_=0
+                )
+            )
+        )
+        await get_database().execute(query)
+
 
 class MeetingConsentController:
     async def get_by_meeting_id(self, meeting_id: str) -> list[MeetingConsent]:
@@ -214,10 +302,9 @@ class MeetingConsentController:
         result = await get_database().fetch_one(query)
         if result is None:
             return None
-        return MeetingConsent(**result) if result else None
+        return MeetingConsent(**result)
 
     async def upsert(self, consent: MeetingConsent) -> MeetingConsent:
-        """Create new consent or update existing one for authenticated users"""
         if consent.user_id:
             # For authenticated users, check if consent already exists
             # not transactional but we're ok with that; the consents ain't deleted anyways

server/reflector/db/recordings.py

@@ -21,6 +21,7 @@ recordings = sa.Table(
         server_default="pending",
     ),
     sa.Column("meeting_id", sa.String),
+    sa.Column("track_keys", sa.JSON, nullable=True),
     sa.Index("idx_recording_meeting_id", "meeting_id"),
 )
 
@@ -28,10 +29,20 @@ recordings = sa.Table(
 class Recording(BaseModel):
     id: str = Field(default_factory=generate_uuid4)
     bucket_name: str
+    # for single-track
     object_key: str
     recorded_at: datetime
     status: Literal["pending", "processing", "completed", "failed"] = "pending"
     meeting_id: str | None = None
+    # for multitrack reprocessing
+    # track_keys can be empty list [] if recording finished but no audio was captured (silence/muted)
+    # None means not a multitrack recording, [] means multitrack with no tracks
+    track_keys: list[str] | None = None
+
+    @property
+    def is_multitrack(self) -> bool:
+        """True if recording has separate audio tracks (1+ tracks counts as multitrack)."""
+        return self.track_keys is not None and len(self.track_keys) > 0
 
 
 class RecordingController:
@@ -40,12 +51,14 @@ class RecordingController:
         await get_database().execute(query)
         return recording
 
-    async def get_by_id(self, id: str) -> Recording:
+    async def get_by_id(self, id: str) -> Recording | None:
         query = recordings.select().where(recordings.c.id == id)
         result = await get_database().fetch_one(query)
         return Recording(**result) if result else None
 
-    async def get_by_object_key(self, bucket_name: str, object_key: str) -> Recording:
+    async def get_by_object_key(
+        self, bucket_name: str, object_key: str
+    ) -> Recording | None:
         query = recordings.select().where(
             recordings.c.bucket_name == bucket_name,
             recordings.c.object_key == object_key,
@@ -57,5 +70,14 @@ class RecordingController:
         query = recordings.delete().where(recordings.c.id == id)
         await get_database().execute(query)
 
+    # no check for existence
+    async def get_by_ids(self, recording_ids: list[str]) -> list[Recording]:
+        if not recording_ids:
+            return []
+
+        query = recordings.select().where(recordings.c.id.in_(recording_ids))
+        results = await get_database().fetch_all(query)
+        return [Recording(**row) for row in results]
+
 
 recordings_controller = RecordingController()

server/reflector/db/rooms.py

@@ -1,3 +1,4 @@
+import secrets
 from datetime import datetime, timezone
 from sqlite3 import IntegrityError
 from typing import Literal
@@ -8,6 +9,8 @@ from pydantic import BaseModel, Field
 from sqlalchemy.sql import false, or_
 
 from reflector.db import get_database, metadata
+from reflector.schemas.platform import Platform
+from reflector.settings import settings
 from reflector.utils import generate_uuid4
 
 rooms = sqlalchemy.Table(
@@ -40,7 +43,22 @@ rooms = sqlalchemy.Table(
     sqlalchemy.Column(
         "is_shared", sqlalchemy.Boolean, nullable=False, server_default=false()
     ),
+    sqlalchemy.Column("webhook_url", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column("webhook_secret", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column("ics_url", sqlalchemy.Text),
+    sqlalchemy.Column("ics_fetch_interval", sqlalchemy.Integer, server_default="300"),
+    sqlalchemy.Column(
+        "ics_enabled", sqlalchemy.Boolean, nullable=False, server_default=false()
+    ),
+    sqlalchemy.Column("ics_last_sync", sqlalchemy.DateTime(timezone=True)),
+    sqlalchemy.Column("ics_last_etag", sqlalchemy.Text),
+    sqlalchemy.Column(
+        "platform",
+        sqlalchemy.String,
+        nullable=False,
+    ),
     sqlalchemy.Index("idx_room_is_shared", "is_shared"),
+    sqlalchemy.Index("idx_room_ics_enabled", "ics_enabled"),
 )
 
 
@@ -55,10 +73,18 @@ class Room(BaseModel):
     is_locked: bool = False
     room_mode: Literal["normal", "group"] = "normal"
     recording_type: Literal["none", "local", "cloud"] = "cloud"
-    recording_trigger: Literal[
+    recording_trigger: Literal[  # whereby-specific
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     is_shared: bool = False
+    webhook_url: str | None = None
+    webhook_secret: str | None = None
+    ics_url: str | None = None
+    ics_fetch_interval: int = 300
+    ics_enabled: bool = False
+    ics_last_sync: datetime | None = None
+    ics_last_etag: str | None = None
+    platform: Platform = Field(default_factory=lambda: settings.DEFAULT_VIDEO_PLATFORM)
 
 
 class RoomController:
@@ -107,22 +133,39 @@ class RoomController:
         recording_type: str,
         recording_trigger: str,
         is_shared: bool,
+        webhook_url: str = "",
+        webhook_secret: str = "",
+        ics_url: str | None = None,
+        ics_fetch_interval: int = 300,
+        ics_enabled: bool = False,
+        platform: Platform = settings.DEFAULT_VIDEO_PLATFORM,
     ):
         """
         Add a new room
         """
-        room = Room(
-            name=name,
-            user_id=user_id,
-            zulip_auto_post=zulip_auto_post,
-            zulip_stream=zulip_stream,
-            zulip_topic=zulip_topic,
-            is_locked=is_locked,
-            room_mode=room_mode,
-            recording_type=recording_type,
-            recording_trigger=recording_trigger,
-            is_shared=is_shared,
-        )
+        if webhook_url and not webhook_secret:
+            webhook_secret = secrets.token_urlsafe(32)
+
+        room_data = {
+            "name": name,
+            "user_id": user_id,
+            "zulip_auto_post": zulip_auto_post,
+            "zulip_stream": zulip_stream,
+            "zulip_topic": zulip_topic,
+            "is_locked": is_locked,
+            "room_mode": room_mode,
+            "recording_type": recording_type,
+            "recording_trigger": recording_trigger,
+            "is_shared": is_shared,
+            "webhook_url": webhook_url,
+            "webhook_secret": webhook_secret,
+            "ics_url": ics_url,
+            "ics_fetch_interval": ics_fetch_interval,
+            "ics_enabled": ics_enabled,
+            "platform": platform,
+        }
+
+        room = Room(**room_data)
         query = rooms.insert().values(**room.model_dump())
         try:
             await get_database().execute(query)
@@ -134,6 +177,9 @@ class RoomController:
         """
         Update a room fields with key/values in values
         """
+        if values.get("webhook_url") and not values.get("webhook_secret"):
+            values["webhook_secret"] = secrets.token_urlsafe(32)
+
         query = rooms.update().where(rooms.c.id == room.id).values(**values)
         try:
             await get_database().execute(query)
@@ -183,6 +229,13 @@ class RoomController:
 
         return room
 
+    async def get_ics_enabled(self) -> list[Room]:
+        query = rooms.select().where(
+            rooms.c.ics_enabled == True, rooms.c.ics_url != None
+        )
+        results = await get_database().fetch_all(query)
+        return [Room(**result) for result in results]
+
     async def remove_by_id(
         self,
         room_id: str,

server/reflector/db/search.py

@@ -8,12 +8,14 @@ from typing import Annotated, Any, Dict, Iterator
 
 import sqlalchemy
 import webvtt
+from databases.interfaces import Record as DbRecord
 from fastapi import HTTPException
 from pydantic import (
     BaseModel,
     Field,
     NonNegativeFloat,
     NonNegativeInt,
+    TypeAdapter,
     ValidationError,
     constr,
     field_serializer,
@@ -21,9 +23,10 @@ from pydantic import (
 
 from reflector.db import get_database
 from reflector.db.rooms import rooms
-from reflector.db.transcripts import SourceKind, transcripts
+from reflector.db.transcripts import SourceKind, TranscriptStatus, transcripts
 from reflector.db.utils import is_postgresql
 from reflector.logger import logger
+from reflector.utils.string import NonEmptyString, try_parse_non_empty_string
 
 DEFAULT_SEARCH_LIMIT = 20
 SNIPPET_CONTEXT_LENGTH = 50  # Characters before/after match to include
@@ -31,12 +34,13 @@ DEFAULT_SNIPPET_MAX_LENGTH = NonNegativeInt(150)
 DEFAULT_MAX_SNIPPETS = NonNegativeInt(3)
 LONG_SUMMARY_MAX_SNIPPETS = 2
 
-SearchQueryBase = constr(min_length=0, strip_whitespace=True)
+SearchQueryBase = constr(min_length=1, strip_whitespace=True)
 SearchLimitBase = Annotated[int, Field(ge=1, le=100)]
 SearchOffsetBase = Annotated[int, Field(ge=0)]
 SearchTotalBase = Annotated[int, Field(ge=0)]
 
 SearchQuery = Annotated[SearchQueryBase, Field(description="Search query text")]
+search_query_adapter = TypeAdapter(SearchQuery)
 SearchLimit = Annotated[SearchLimitBase, Field(description="Results per page")]
 SearchOffset = Annotated[
     SearchOffsetBase, Field(description="Number of results to skip")
@@ -88,7 +92,7 @@ class WebVTTProcessor:
     @staticmethod
     def generate_snippets(
         webvtt_content: WebVTTContent,
-        query: str,
+        query: SearchQuery,
         max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from WebVTT content."""
@@ -125,12 +129,14 @@ class SnippetCandidate:
 class SearchParameters(BaseModel):
     """Validated search parameters for full-text search."""
 
-    query_text: SearchQuery
+    query_text: SearchQuery | None = None
     limit: SearchLimit = DEFAULT_SEARCH_LIMIT
     offset: SearchOffset = 0
     user_id: str | None = None
     room_id: str | None = None
     source_kind: SourceKind | None = None
+    from_datetime: datetime | None = None
+    to_datetime: datetime | None = None
 
 
 class SearchResultDB(BaseModel):
@@ -157,7 +163,7 @@ class SearchResult(BaseModel):
     room_name: str | None = None
     source_kind: SourceKind
     created_at: datetime
-    status: str = Field(..., min_length=1)
+    status: TranscriptStatus = Field(..., min_length=1)
     rank: float = Field(..., ge=0, le=1)
     duration: NonNegativeFloat | None = Field(..., description="Duration in seconds")
     search_snippets: list[str] = Field(
@@ -199,15 +205,13 @@ class SnippetGenerator:
             prev_start = start
 
     @staticmethod
-    def count_matches(text: str, query: str) -> NonNegativeInt:
+    def count_matches(text: str, query: SearchQuery) -> NonNegativeInt:
         """Count total number of matches for a query in text."""
         ZERO = NonNegativeInt(0)
         if not text:
             logger.warning("Empty text for search query in count_matches")
             return ZERO
-        if not query:
-            logger.warning("Empty query for search text in count_matches")
-            return ZERO
+        assert query is not None
         return NonNegativeInt(
             sum(1 for _ in SnippetGenerator.find_all_matches(text, query))
         )
@@ -243,13 +247,14 @@ class SnippetGenerator:
     @staticmethod
     def generate(
         text: str,
-        query: str,
+        query: SearchQuery,
         max_length: NonNegativeInt = DEFAULT_SNIPPET_MAX_LENGTH,
         max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from text."""
-        if not text or not query:
-            logger.warning("Empty text or query for generate_snippets")
+        assert query is not None
+        if not text:
+            logger.warning("Empty text for generate_snippets")
             return []
 
         candidates = (
@@ -270,7 +275,7 @@ class SnippetGenerator:
     @staticmethod
     def from_summary(
         summary: str,
-        query: str,
+        query: SearchQuery,
         max_snippets: NonNegativeInt = LONG_SUMMARY_MAX_SNIPPETS,
     ) -> list[str]:
         """Generate snippets from summary text."""
@@ -278,9 +283,9 @@ class SnippetGenerator:
 
     @staticmethod
     def combine_sources(
-        summary: str | None,
+        summary: NonEmptyString | None,
         webvtt: WebVTTContent | None,
-        query: str,
+        query: SearchQuery,
         max_total: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
     ) -> tuple[list[str], NonNegativeInt]:
         """Combine snippets from multiple sources and return total match count.
@@ -289,6 +294,11 @@ class SnippetGenerator:
 
         snippets can be empty for real in case of e.g. title match
         """
+
+        assert (
+            summary is not None or webvtt is not None
+        ), "At least one source must be present"
+
         webvtt_matches = 0
         summary_matches = 0
 
@@ -355,8 +365,8 @@ class SearchController:
                 else_=rooms.c.name,
             ).label("room_name"),
         ]
-
-        if params.query_text:
+        search_query = None
+        if params.query_text is not None:
             search_query = sqlalchemy.func.websearch_to_tsquery(
                 "english", params.query_text
             )
@@ -373,7 +383,9 @@ class SearchController:
             transcripts.join(rooms, transcripts.c.room_id == rooms.c.id, isouter=True)
         )
 
-        if params.query_text:
+        if params.query_text is not None:
+            # because already initialized based on params.query_text presence above
+            assert search_query is not None
             base_query = base_query.where(
                 transcripts.c.search_vector_en.op("@@")(search_query)
             )
@@ -392,8 +404,16 @@ class SearchController:
             base_query = base_query.where(
                 transcripts.c.source_kind == params.source_kind
             )
+        if params.from_datetime:
+            base_query = base_query.where(
+                transcripts.c.created_at >= params.from_datetime
+            )
+        if params.to_datetime:
+            base_query = base_query.where(
+                transcripts.c.created_at <= params.to_datetime
+            )
 
-        if params.query_text:
+        if params.query_text is not None:
             order_by = sqlalchemy.desc(sqlalchemy.text("rank"))
         else:
             order_by = sqlalchemy.desc(transcripts.c.created_at)
@@ -407,19 +427,29 @@ class SearchController:
         )
         total = await get_database().fetch_val(count_query)
 
-        def _process_result(r) -> SearchResult:
+        def _process_result(r: DbRecord) -> SearchResult:
             r_dict: Dict[str, Any] = dict(r)
+
             webvtt_raw: str | None = r_dict.pop("webvtt", None)
+            webvtt: WebVTTContent | None
             if webvtt_raw:
                 webvtt = WebVTTProcessor.parse(webvtt_raw)
             else:
                 webvtt = None
-            long_summary: str | None = r_dict.pop("long_summary", None)
+
+            long_summary_r: str | None = r_dict.pop("long_summary", None)
+            long_summary: NonEmptyString = try_parse_non_empty_string(long_summary_r)
             room_name: str | None = r_dict.pop("room_name", None)
             db_result = SearchResultDB.model_validate(r_dict)
 
-            snippets, total_match_count = SnippetGenerator.combine_sources(
-                long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+            at_least_one_source = webvtt is not None or long_summary is not None
+            has_query = params.query_text is not None
+            snippets, total_match_count = (
+                SnippetGenerator.combine_sources(
+                    long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+                )
+                if has_query and at_least_one_source
+                else ([], 0)
             )
 
             return SearchResult(

server/reflector/db/transcripts.py

@@ -21,7 +21,7 @@ from reflector.db.utils import is_postgresql
 from reflector.logger import logger
 from reflector.processors.types import Word as ProcessorWord
 from reflector.settings import settings
-from reflector.storage import get_recordings_storage, get_transcripts_storage
+from reflector.storage import get_transcripts_storage
 from reflector.utils import generate_uuid4
 from reflector.utils.webvtt import topics_to_webvtt
 
@@ -122,6 +122,15 @@ def generate_transcript_name() -> str:
     return f"Transcript {now.strftime('%Y-%m-%d %H:%M:%S')}"
 
 
+TranscriptStatus = Literal[
+    "idle", "uploaded", "recording", "processing", "error", "ended"
+]
+
+
+class StrValue(BaseModel):
+    value: str
+
+
 class AudioWaveform(BaseModel):
     data: list[float]
 
@@ -177,6 +186,7 @@ class TranscriptParticipant(BaseModel):
     id: str = Field(default_factory=generate_uuid4)
     speaker: int | None
     name: str
+    user_id: str | None = None
 
 
 class Transcript(BaseModel):
@@ -185,7 +195,7 @@ class Transcript(BaseModel):
     id: str = Field(default_factory=generate_uuid4)
     user_id: str | None = None
     name: str = Field(default_factory=generate_transcript_name)
-    status: str = "idle"
+    status: TranscriptStatus = "idle"
     duration: float = 0
     created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     title: str | None = None
@@ -614,7 +624,9 @@ class TranscriptController:
                 )
                 if recording:
                     try:
-                        await get_recordings_storage().delete_file(recording.object_key)
+                        await get_transcripts_storage().delete_file(
+                            recording.object_key, bucket=recording.bucket_name
+                        )
                     except Exception as e:
                         logger.warning(
                             "Failed to delete recording object from S3",
@@ -638,6 +650,19 @@ class TranscriptController:
         query = transcripts.delete().where(transcripts.c.recording_id == recording_id)
         await get_database().execute(query)
 
+    @staticmethod
+    def user_can_mutate(transcript: Transcript, user_id: str | None) -> bool:
+        """
+        Returns True if the given user is allowed to modify the transcript.
+
+        Policy:
+        - Anonymous transcripts (user_id is None) cannot be modified via API
+        - Only the owner (matching user_id) can modify their transcript
+        """
+        if transcript.user_id is None:
+            return False
+        return user_id and transcript.user_id == user_id
+
     @asynccontextmanager
     async def transaction(self):
         """
@@ -703,11 +728,13 @@ class TranscriptController:
         """
         Download audio from storage
         """
-        transcript.audio_mp3_filename.write_bytes(
-            await get_transcripts_storage().get_file(
-                transcript.storage_audio_path,
-            )
-        )
+        storage = get_transcripts_storage()
+        try:
+            with open(transcript.audio_mp3_filename, "wb") as f:
+                await storage.stream_to_fileobj(transcript.storage_audio_path, f)
+        except Exception:
+            transcript.audio_mp3_filename.unlink(missing_ok=True)
+            raise
 
     async def upsert_participant(
         self,
@@ -732,5 +759,27 @@ class TranscriptController:
         transcript.delete_participant(participant_id)
         await self.update(transcript, {"participants": transcript.participants_dump()})
 
+    async def set_status(
+        self, transcript_id: str, status: TranscriptStatus
+    ) -> TranscriptEvent | None:
+        """
+        Update the status of a transcript
+
+        Will add an event STATUS + update the status field of transcript
+        """
+        async with self.transaction():
+            transcript = await self.get_by_id(transcript_id)
+            if not transcript:
+                raise Exception(f"Transcript {transcript_id} not found")
+            if transcript.status == status:
+                return
+            resp = await self.append_event(
+                transcript=transcript,
+                event="STATUS",
+                data=StrValue(value=status),
+            )
+            await self.update(transcript, {"status": status})
+        return resp
+
 
 transcripts_controller = TranscriptController()

server/reflector/db/user_api_keys.py

@@ -0,0 +1,91 @@
+import hmac
+import secrets
+from datetime import datetime, timezone
+from hashlib import sha256
+
+import sqlalchemy
+from pydantic import BaseModel, Field
+
+from reflector.db import get_database, metadata
+from reflector.settings import settings
+from reflector.utils import generate_uuid4
+from reflector.utils.string import NonEmptyString
+
+user_api_keys = sqlalchemy.Table(
+    "user_api_key",
+    metadata,
+    sqlalchemy.Column("id", sqlalchemy.String, primary_key=True),
+    sqlalchemy.Column("user_id", sqlalchemy.String, nullable=False),
+    sqlalchemy.Column("key_hash", sqlalchemy.String, nullable=False),
+    sqlalchemy.Column("name", sqlalchemy.String, nullable=True),
+    sqlalchemy.Column("created_at", sqlalchemy.DateTime(timezone=True), nullable=False),
+    sqlalchemy.Index("idx_user_api_key_hash", "key_hash", unique=True),
+    sqlalchemy.Index("idx_user_api_key_user_id", "user_id"),
+)
+
+
+class UserApiKey(BaseModel):
+    id: NonEmptyString = Field(default_factory=generate_uuid4)
+    user_id: NonEmptyString
+    key_hash: NonEmptyString
+    name: NonEmptyString | None = None
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class UserApiKeyController:
+    @staticmethod
+    def generate_key() -> NonEmptyString:
+        return secrets.token_urlsafe(48)
+
+    @staticmethod
+    def hash_key(key: NonEmptyString) -> str:
+        return hmac.new(
+            settings.SECRET_KEY.encode(), key.encode(), digestmod=sha256
+        ).hexdigest()
+
+    @classmethod
+    async def create_key(
+        cls,
+        user_id: NonEmptyString,
+        name: NonEmptyString | None = None,
+    ) -> tuple[UserApiKey, NonEmptyString]:
+        plaintext = cls.generate_key()
+        api_key = UserApiKey(
+            user_id=user_id,
+            key_hash=cls.hash_key(plaintext),
+            name=name,
+        )
+        query = user_api_keys.insert().values(**api_key.model_dump())
+        await get_database().execute(query)
+        return api_key, plaintext
+
+    @classmethod
+    async def verify_key(cls, plaintext_key: NonEmptyString) -> UserApiKey | None:
+        key_hash = cls.hash_key(plaintext_key)
+        query = user_api_keys.select().where(
+            user_api_keys.c.key_hash == key_hash,
+        )
+        result = await get_database().fetch_one(query)
+        return UserApiKey(**result) if result else None
+
+    @staticmethod
+    async def list_by_user_id(user_id: NonEmptyString) -> list[UserApiKey]:
+        query = (
+            user_api_keys.select()
+            .where(user_api_keys.c.user_id == user_id)
+            .order_by(user_api_keys.c.created_at.desc())
+        )
+        results = await get_database().fetch_all(query)
+        return [UserApiKey(**r) for r in results]
+
+    @staticmethod
+    async def delete_key(key_id: NonEmptyString, user_id: NonEmptyString) -> bool:
+        query = user_api_keys.delete().where(
+            (user_api_keys.c.id == key_id) & (user_api_keys.c.user_id == user_id)
+        )
+        result = await get_database().execute(query)
+        # asyncpg returns None for DELETE, consider it success if no exception
+        return result is None or result > 0
+
+
+user_api_keys_controller = UserApiKeyController()

server/reflector/db/users.py

@@ -0,0 +1,92 @@
+"""User table for storing Authentik user information."""
+
+from datetime import datetime, timezone
+
+import sqlalchemy
+from pydantic import BaseModel, Field
+
+from reflector.db import get_database, metadata
+from reflector.utils import generate_uuid4
+from reflector.utils.string import NonEmptyString
+
+users = sqlalchemy.Table(
+    "user",
+    metadata,
+    sqlalchemy.Column("id", sqlalchemy.String, primary_key=True),
+    sqlalchemy.Column("email", sqlalchemy.String, nullable=False),
+    sqlalchemy.Column("authentik_uid", sqlalchemy.String, nullable=False),
+    sqlalchemy.Column("created_at", sqlalchemy.DateTime(timezone=True), nullable=False),
+    sqlalchemy.Column("updated_at", sqlalchemy.DateTime(timezone=True), nullable=False),
+    sqlalchemy.Index("idx_user_authentik_uid", "authentik_uid", unique=True),
+    sqlalchemy.Index("idx_user_email", "email", unique=False),
+)
+
+
+class User(BaseModel):
+    id: NonEmptyString = Field(default_factory=generate_uuid4)
+    email: NonEmptyString
+    authentik_uid: NonEmptyString
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class UserController:
+    @staticmethod
+    async def get_by_id(user_id: NonEmptyString) -> User | None:
+        query = users.select().where(users.c.id == user_id)
+        result = await get_database().fetch_one(query)
+        return User(**result) if result else None
+
+    @staticmethod
+    async def get_by_authentik_uid(authentik_uid: NonEmptyString) -> User | None:
+        query = users.select().where(users.c.authentik_uid == authentik_uid)
+        result = await get_database().fetch_one(query)
+        return User(**result) if result else None
+
+    @staticmethod
+    async def get_by_email(email: NonEmptyString) -> User | None:
+        query = users.select().where(users.c.email == email)
+        result = await get_database().fetch_one(query)
+        return User(**result) if result else None
+
+    @staticmethod
+    async def create_or_update(
+        id: NonEmptyString, authentik_uid: NonEmptyString, email: NonEmptyString
+    ) -> User:
+        existing = await UserController.get_by_authentik_uid(authentik_uid)
+        now = datetime.now(timezone.utc)
+
+        if existing:
+            query = (
+                users.update()
+                .where(users.c.authentik_uid == authentik_uid)
+                .values(email=email, updated_at=now)
+            )
+            await get_database().execute(query)
+            return User(
+                id=existing.id,
+                authentik_uid=authentik_uid,
+                email=email,
+                created_at=existing.created_at,
+                updated_at=now,
+            )
+        else:
+            user = User(
+                id=id,
+                authentik_uid=authentik_uid,
+                email=email,
+                created_at=now,
+                updated_at=now,
+            )
+            query = users.insert().values(**user.model_dump())
+            await get_database().execute(query)
+            return user
+
+    @staticmethod
+    async def list_all() -> list[User]:
+        query = users.select().order_by(users.c.created_at.desc())
+        results = await get_database().fetch_all(query)
+        return [User(**r) for r in results]
+
+
+user_controller = UserController()

server/reflector/llm.py

@@ -1,3 +1,4 @@
+import logging
 from typing import Type, TypeVar
 
 from llama_index.core import Settings
@@ -5,7 +6,7 @@ from llama_index.core.output_parsers import PydanticOutputParser
 from llama_index.core.program import LLMTextCompletionProgram
 from llama_index.core.response_synthesizers import TreeSummarize
 from llama_index.llms.openai_like import OpenAILike
-from pydantic import BaseModel
+from pydantic import BaseModel, ValidationError
 
 T = TypeVar("T", bound=BaseModel)
 
@@ -61,6 +62,8 @@ class LLM:
         tone_name: str | None = None,
     ) -> T:
         """Get structured output from LLM for non-function-calling models"""
+        logger = logging.getLogger(__name__)
+
         summarizer = TreeSummarize(verbose=True)
         response = await summarizer.aget_response(prompt, texts, tone_name=tone_name)
 
@@ -76,8 +79,25 @@ class LLM:
             "Please structure the above information in the following JSON format:"
         )
 
-        output = await program.acall(
-            analysis=str(response), format_instructions=format_instructions
-        )
+        try:
+            output = await program.acall(
+                analysis=str(response), format_instructions=format_instructions
+            )
+        except ValidationError as e:
+            # Extract the raw JSON from the error details
+            errors = e.errors()
+            if errors and "input" in errors[0]:
+                raw_json = errors[0]["input"]
+                logger.error(
+                    f"JSON validation failed for {output_cls.__name__}. "
+                    f"Full raw JSON output:\n{raw_json}\n"
+                    f"Validation errors: {errors}"
+                )
+            else:
+                logger.error(
+                    f"JSON validation failed for {output_cls.__name__}. "
+                    f"Validation errors: {errors}"
+                )
+            raise
 
         return output

server/reflector/pipelines/__init__.py

@@ -0,0 +1 @@
+"""Pipeline modules for audio processing."""

server/reflector/pipelines/main_file_pipeline.py

@@ -7,29 +7,34 @@ Uses parallel processing for transcription, diarization, and waveform generation
 """
 
 import asyncio
+import uuid
 from pathlib import Path
 
 import av
 import structlog
-from celery import shared_task
+from celery import chain, shared_task
 
+from reflector.asynctask import asynctask
+from reflector.db.rooms import rooms_controller
 from reflector.db.transcripts import (
+    SourceKind,
     Transcript,
+    TranscriptStatus,
     transcripts_controller,
 )
 from reflector.logger import logger
-from reflector.pipelines.main_live_pipeline import PipelineMainBase, asynctask
-from reflector.processors import (
-    AudioFileWriterProcessor,
-    TranscriptFinalSummaryProcessor,
-    TranscriptFinalTitleProcessor,
-    TranscriptTopicDetectorProcessor,
+from reflector.pipelines import topic_processing
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    broadcast_to_sockets,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
 )
+from reflector.pipelines.transcription_helpers import transcribe_file_with_processor
+from reflector.processors import AudioFileWriterProcessor
 from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
 from reflector.processors.file_diarization import FileDiarizationInput
 from reflector.processors.file_diarization_auto import FileDiarizationAutoProcessor
-from reflector.processors.file_transcript import FileTranscriptInput
-from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
 from reflector.processors.transcript_diarization_assembler import (
     TranscriptDiarizationAssemblerInput,
     TranscriptDiarizationAssemblerProcessor,
@@ -43,19 +48,7 @@ from reflector.processors.types import (
 )
 from reflector.settings import settings
 from reflector.storage import get_transcripts_storage
-
-
-class EmptyPipeline:
-    """Empty pipeline for processors that need a pipeline reference"""
-
-    def __init__(self, logger: structlog.BoundLogger):
-        self.logger = logger
-
-    def get_pref(self, k, d=None):
-        return d
-
-    async def emit(self, event):
-        pass
+from reflector.worker.webhook import send_transcript_webhook
 
 
 class PipelineMainFile(PipelineMainBase):
@@ -70,7 +63,7 @@ class PipelineMainFile(PipelineMainBase):
     def __init__(self, transcript_id: str):
         super().__init__(transcript_id=transcript_id)
         self.logger = logger.bind(transcript_id=self.transcript_id)
-        self.empty_pipeline = EmptyPipeline(logger=self.logger)
+        self.empty_pipeline = topic_processing.EmptyPipeline(logger=self.logger)
 
     def _handle_gather_exceptions(self, results: list, operation: str) -> None:
         """Handle exceptions from asyncio.gather with return_exceptions=True"""
@@ -83,12 +76,27 @@ class PipelineMainFile(PipelineMainBase):
                 exc_info=result,
             )
 
+    @broadcast_to_sockets
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
     async def process(self, file_path: Path):
         """Main entry point for file processing"""
         self.logger.info(f"Starting file pipeline for {file_path}")
 
         transcript = await self.get_transcript()
 
+        # Clear transcript as we're going to regenerate everything
+        async with self.transaction():
+            await transcripts_controller.update(
+                transcript,
+                {
+                    "events": [],
+                    "topics": [],
+                },
+            )
+
         # Extract audio and write to transcript location
         audio_path = await self.extract_and_write_audio(file_path, transcript)
 
@@ -105,6 +113,8 @@ class PipelineMainFile(PipelineMainBase):
 
         self.logger.info("File pipeline complete")
 
+        await self.set_status(transcript.id, "ended")
+
     async def extract_and_write_audio(
         self, file_path: Path, transcript: Transcript
     ) -> Path:
@@ -234,24 +244,7 @@ class PipelineMainFile(PipelineMainBase):
 
     async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
         """Transcribe complete file"""
-        processor = FileTranscriptAutoProcessor()
-        input_data = FileTranscriptInput(audio_url=audio_url, language=language)
-
-        # Store result for retrieval
-        result: TranscriptType | None = None
-
-        async def capture_result(transcript):
-            nonlocal result
-            result = transcript
-
-        processor.on(capture_result)
-        await processor.push(input_data)
-        await processor.flush()
-
-        if not result:
-            raise ValueError("No transcript captured")
-
-        return result
+        return await transcribe_file_with_processor(audio_url, language)
 
     async def diarize_file(self, audio_url: str) -> list[DiarizationSegment] | None:
         """Get diarization for file"""
@@ -294,63 +287,53 @@ class PipelineMainFile(PipelineMainBase):
     async def detect_topics(
         self, transcript: TranscriptType, target_language: str
     ) -> list[TitleSummary]:
-        """Detect topics from complete transcript"""
-        chunk_size = 300
-        topics: list[TitleSummary] = []
-
-        async def on_topic(topic: TitleSummary):
-            topics.append(topic)
-            return await self.on_topic(topic)
-
-        topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
-        topic_detector.set_pipeline(self.empty_pipeline)
-
-        for i in range(0, len(transcript.words), chunk_size):
-            chunk_words = transcript.words[i : i + chunk_size]
-            if not chunk_words:
-                continue
-
-            chunk_transcript = TranscriptType(
-                words=chunk_words, translation=transcript.translation
-            )
-
-            await topic_detector.push(chunk_transcript)
-
-        await topic_detector.flush()
-        return topics
+        return await topic_processing.detect_topics(
+            transcript,
+            target_language,
+            on_topic_callback=self.on_topic,
+            empty_pipeline=self.empty_pipeline,
+        )
 
     async def generate_title(self, topics: list[TitleSummary]):
-        """Generate title from topics"""
-        if not topics:
-            self.logger.warning("No topics for title generation")
-            return
-
-        processor = TranscriptFinalTitleProcessor(callback=self.on_title)
-        processor.set_pipeline(self.empty_pipeline)
-
-        for topic in topics:
-            await processor.push(topic)
-
-        await processor.flush()
+        return await topic_processing.generate_title(
+            topics,
+            on_title_callback=self.on_title,
+            empty_pipeline=self.empty_pipeline,
+            logger=self.logger,
+        )
 
     async def generate_summaries(self, topics: list[TitleSummary]):
-        """Generate long and short summaries from topics"""
-        if not topics:
-            self.logger.warning("No topics for summary generation")
-            return
-
         transcript = await self.get_transcript()
-        processor = TranscriptFinalSummaryProcessor(
-            transcript=transcript,
-            callback=self.on_long_summary,
-            on_short_summary=self.on_short_summary,
+        return await topic_processing.generate_summaries(
+            topics,
+            transcript,
+            on_long_summary_callback=self.on_long_summary,
+            on_short_summary_callback=self.on_short_summary,
+            empty_pipeline=self.empty_pipeline,
+            logger=self.logger,
         )
-        processor.set_pipeline(self.empty_pipeline)
 
-        for topic in topics:
-            await processor.push(topic)
 
-        await processor.flush()
+@shared_task
+@asynctask
+async def task_send_webhook_if_needed(*, transcript_id: str):
+    """Send webhook if this is a room recording with webhook configured"""
+    transcript = await transcripts_controller.get_by_id(transcript_id)
+    if not transcript:
+        return
+
+    if transcript.source_kind == SourceKind.ROOM and transcript.room_id:
+        room = await rooms_controller.get_by_id(transcript.room_id)
+        if room and room.webhook_url:
+            logger.info(
+                "Dispatching webhook",
+                transcript_id=transcript_id,
+                room_id=room.id,
+                webhook_url=room.webhook_url,
+            )
+            send_transcript_webhook.delay(
+                transcript_id, room.id, event_id=uuid.uuid4().hex
+            )
 
 
 @shared_task
@@ -362,14 +345,33 @@ async def task_pipeline_file_process(*, transcript_id: str):
     if not transcript:
         raise Exception(f"Transcript {transcript_id} not found")
 
-    # Find the file to process
-    audio_file = next(transcript.data_path.glob("upload.*"), None)
-    if not audio_file:
-        audio_file = next(transcript.data_path.glob("audio.*"), None)
-
-    if not audio_file:
-        raise Exception("No audio file found to process")
-
-    # Run file pipeline
     pipeline = PipelineMainFile(transcript_id=transcript_id)
-    await pipeline.process(audio_file)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+
+        # Find the file to process
+        audio_file = next(transcript.data_path.glob("upload.*"), None)
+        if not audio_file:
+            audio_file = next(transcript.data_path.glob("audio.*"), None)
+
+        if not audio_file:
+            raise Exception("No audio file found to process")
+
+        await pipeline.process(audio_file)
+
+    except Exception as e:
+        logger.error(
+            f"File pipeline failed for transcript {transcript_id}: {type(e).__name__}: {str(e)}",
+            exc_info=True,
+            transcript_id=transcript_id,
+        )
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    # Run post-processing chain: consent cleanup -> zulip -> webhook
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/main_live_pipeline.py

@@ -17,12 +17,11 @@ from contextlib import asynccontextmanager
 from typing import Generic
 
 import av
-import boto3
 from celery import chord, current_task, group, shared_task
 from pydantic import BaseModel
 from structlog import BoundLogger as Logger
 
-from reflector.db import get_database
+from reflector.asynctask import asynctask
 from reflector.db.meetings import meeting_consent_controller, meetings_controller
 from reflector.db.recordings import recordings_controller
 from reflector.db.rooms import rooms_controller
@@ -32,6 +31,7 @@ from reflector.db.transcripts import (
     TranscriptFinalLongSummary,
     TranscriptFinalShortSummary,
     TranscriptFinalTitle,
+    TranscriptStatus,
     TranscriptText,
     TranscriptTopic,
     TranscriptWaveform,
@@ -40,8 +40,9 @@ from reflector.db.transcripts import (
 from reflector.logger import logger
 from reflector.pipelines.runner import PipelineMessage, PipelineRunner
 from reflector.processors import (
-    AudioChunkerProcessor,
+    AudioChunkerAutoProcessor,
     AudioDiarizationAutoProcessor,
+    AudioDownscaleProcessor,
     AudioFileWriterProcessor,
     AudioMergeProcessor,
     AudioTranscriptAutoProcessor,
@@ -68,29 +69,6 @@ from reflector.zulip import (
 )
 
 
-def asynctask(f):
-    @functools.wraps(f)
-    def wrapper(*args, **kwargs):
-        async def run_with_db():
-            database = get_database()
-            await database.connect()
-            try:
-                return await f(*args, **kwargs)
-            finally:
-                await database.disconnect()
-
-        coro = run_with_db()
-        try:
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            loop = None
-        if loop and loop.is_running():
-            return loop.run_until_complete(coro)
-        return asyncio.run(coro)
-
-    return wrapper
-
-
 def broadcast_to_sockets(func):
     """
     Decorator to broadcast transcript event to websockets
@@ -106,6 +84,20 @@ def broadcast_to_sockets(func):
             message=resp.model_dump(mode="json"),
         )
 
+        transcript = await transcripts_controller.get_by_id(self.transcript_id)
+        if transcript and transcript.user_id:
+            # Emit only relevant events to the user room to avoid noisy updates.
+            # Allowed: STATUS, FINAL_TITLE, DURATION. All are prefixed with TRANSCRIPT_
+            allowed_user_events = {"STATUS", "FINAL_TITLE", "DURATION"}
+            if resp.event in allowed_user_events:
+                await self.ws_manager.send_json(
+                    room_id=f"user:{transcript.user_id}",
+                    message={
+                        "event": f"TRANSCRIPT_{resp.event}",
+                        "data": {"id": self.transcript_id, **resp.data},
+                    },
+                )
+
     return wrapper
 
 
@@ -187,8 +179,15 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
         ]
 
     @asynccontextmanager
-    async def transaction(self):
+    async def lock_transaction(self):
+        # This lock is to prevent multiple processor starting adding
+        # into event array at the same time
         async with self._lock:
+            yield
+
+    @asynccontextmanager
+    async def transaction(self):
+        async with self.lock_transaction():
             async with transcripts_controller.transaction():
                 yield
 
@@ -197,14 +196,14 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
         # if it's the first part, update the status of the transcript
         # but do not set the ended status yet.
         if isinstance(self, PipelineMainLive):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "started": "recording",
                 "push": "recording",
                 "flush": "processing",
                 "error": "error",
             }
         elif isinstance(self, PipelineMainFinalSummaries):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "push": "processing",
                 "flush": "processing",
                 "error": "error",
@@ -220,22 +219,8 @@ class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]
             return
 
         # when the status of the pipeline changes, update the transcript
-        async with self.transaction():
-            transcript = await self.get_transcript()
-            if status == transcript.status:
-                return
-            resp = await transcripts_controller.append_event(
-                transcript=transcript,
-                event="STATUS",
-                data=StrValue(value=status),
-            )
-            await transcripts_controller.update(
-                transcript,
-                {
-                    "status": status,
-                },
-            )
-            return resp
+        async with self._lock:
+            return await transcripts_controller.set_status(self.transcript_id, status)
 
     @broadcast_to_sockets
     async def on_transcript(self, data):
@@ -365,7 +350,8 @@ class PipelineMainLive(PipelineMainBase):
                 path=transcript.audio_wav_filename,
                 on_duration=self.on_duration,
             ),
-            AudioChunkerProcessor(),
+            AudioDownscaleProcessor(),
+            AudioChunkerAutoProcessor(),
             AudioMergeProcessor(),
             AudioTranscriptAutoProcessor.as_threaded(),
             TranscriptLinerProcessor(),
@@ -597,6 +583,7 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
 
     consent_denied = False
     recording = None
+    meeting = None
     try:
         if transcript.recording_id:
             recording = await recordings_controller.get_by_id(transcript.recording_id)
@@ -607,8 +594,8 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
                         meeting.id
                     )
     except Exception as e:
-        logger.error(f"Failed to get fetch consent: {e}", exc_info=e)
-        consent_denied = True
+        logger.error(f"Failed to fetch consent: {e}", exc_info=e)
+        raise
 
     if not consent_denied:
         logger.info("Consent approved, keeping all files")
@@ -616,25 +603,24 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
 
     logger.info("Consent denied, cleaning up all related audio files")
 
-    if recording and recording.bucket_name and recording.object_key:
-        s3_whereby = boto3.client(
-            "s3",
-            aws_access_key_id=settings.AWS_WHEREBY_ACCESS_KEY_ID,
-            aws_secret_access_key=settings.AWS_WHEREBY_ACCESS_KEY_SECRET,
-        )
-        try:
-            s3_whereby.delete_object(
-                Bucket=recording.bucket_name, Key=recording.object_key
-            )
-            logger.info(
-                f"Deleted original Whereby recording: {recording.bucket_name}/{recording.object_key}"
-            )
-        except Exception as e:
-            logger.error(f"Failed to delete Whereby recording: {e}", exc_info=e)
+    deletion_errors = []
+    if recording and recording.bucket_name:
+        keys_to_delete = []
+        if recording.track_keys:
+            keys_to_delete = recording.track_keys
+        elif recording.object_key:
+            keys_to_delete = [recording.object_key]
+
+        master_storage = get_transcripts_storage()
+        for key in keys_to_delete:
+            try:
+                await master_storage.delete_file(key, bucket=recording.bucket_name)
+                logger.info(f"Deleted recording file: {recording.bucket_name}/{key}")
+            except Exception as e:
+                error_msg = f"Failed to delete {key}: {e}"
+                logger.error(error_msg, exc_info=e)
+                deletion_errors.append(error_msg)
 
-    # non-transactional, files marked for deletion not actually deleted is possible
-    await transcripts_controller.update(transcript, {"audio_deleted": True})
-    # 2. Delete processed audio from transcript storage S3 bucket
     if transcript.audio_location == "storage":
         storage = get_transcripts_storage()
         try:
@@ -643,18 +629,28 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
                 f"Deleted processed audio from storage: {transcript.storage_audio_path}"
             )
         except Exception as e:
-            logger.error(f"Failed to delete processed audio: {e}", exc_info=e)
+            error_msg = f"Failed to delete processed audio: {e}"
+            logger.error(error_msg, exc_info=e)
+            deletion_errors.append(error_msg)
 
-    # 3. Delete local audio files
     try:
         if hasattr(transcript, "audio_mp3_filename") and transcript.audio_mp3_filename:
             transcript.audio_mp3_filename.unlink(missing_ok=True)
         if hasattr(transcript, "audio_wav_filename") and transcript.audio_wav_filename:
             transcript.audio_wav_filename.unlink(missing_ok=True)
     except Exception as e:
-        logger.error(f"Failed to delete local audio files: {e}", exc_info=e)
+        error_msg = f"Failed to delete local audio files: {e}"
+        logger.error(error_msg, exc_info=e)
+        deletion_errors.append(error_msg)
 
-    logger.info("Consent cleanup done")
+    if deletion_errors:
+        logger.warning(
+            f"Consent cleanup completed with {len(deletion_errors)} errors",
+            errors=deletion_errors,
+        )
+    else:
+        await transcripts_controller.update(transcript, {"audio_deleted": True})
+        logger.info("Consent cleanup done - all audio deleted")
 
 
 @get_transcript
@@ -792,7 +788,7 @@ def pipeline_post(*, transcript_id: str):
         chain_final_summaries,
     ) | task_pipeline_post_to_zulip.si(transcript_id=transcript_id)
 
-    chain.delay()
+    return chain.delay()
 
 
 @get_transcript

server/reflector/pipelines/main_multitrack_pipeline.py

@@ -0,0 +1,803 @@
+import asyncio
+import math
+import tempfile
+from fractions import Fraction
+from pathlib import Path
+
+import av
+from av.audio.resampler import AudioResampler
+from celery import chain, shared_task
+
+from reflector.asynctask import asynctask
+from reflector.dailyco_api import MeetingParticipantsResponse
+from reflector.db.transcripts import (
+    Transcript,
+    TranscriptParticipant,
+    TranscriptStatus,
+    TranscriptWaveform,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines import topic_processing
+from reflector.pipelines.main_file_pipeline import task_send_webhook_if_needed
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    broadcast_to_sockets,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
+from reflector.pipelines.transcription_helpers import transcribe_file_with_processor
+from reflector.processors import AudioFileWriterProcessor
+from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
+from reflector.processors.types import TitleSummary
+from reflector.processors.types import Transcript as TranscriptType
+from reflector.settings import settings
+from reflector.storage import Storage, get_transcripts_storage
+from reflector.utils.daily import (
+    filter_cam_audio_tracks,
+    parse_daily_recording_filename,
+)
+from reflector.utils.string import NonEmptyString
+from reflector.video_platforms.factory import create_platform_client
+
+# Audio encoding constants
+OPUS_STANDARD_SAMPLE_RATE = 48000
+OPUS_DEFAULT_BIT_RATE = 128000
+
+# Storage operation constants
+PRESIGNED_URL_EXPIRATION_SECONDS = 7200  # 2 hours
+
+
+class PipelineMainMultitrack(PipelineMainBase):
+    def __init__(self, transcript_id: str):
+        super().__init__(transcript_id=transcript_id)
+        self.logger = logger.bind(transcript_id=self.transcript_id)
+        self.empty_pipeline = topic_processing.EmptyPipeline(logger=self.logger)
+
+    async def pad_track_for_transcription(
+        self,
+        track_url: NonEmptyString,
+        track_idx: int,
+        storage: Storage,
+    ) -> NonEmptyString:
+        """
+        Pad a single track with silence based on stream metadata start_time.
+        Downloads from S3 presigned URL, processes via PyAV using tempfile, uploads to S3.
+        Returns presigned URL of padded track (or original URL if no padding needed).
+
+        Memory usage:
+        - Pattern: fixed_overhead(2-5MB) for PyAV codec/filters
+        - PyAV streams input efficiently (no full download, verified)
+        - Output written to tempfile (disk-based, not memory)
+        - Upload streams from file handle (boto3 chunks, typically 5-10MB)
+
+        Daily.co raw-tracks timing - Two approaches:
+
+            CURRENT APPROACH (PyAV metadata):
+            The WebM stream.start_time field encodes MEETING-RELATIVE timing:
+            - t=0: When Daily.co recording started (first participant joined)
+            - start_time=8.13s: This participant's track began 8.13s after recording started
+            - Purpose: Enables track alignment without external manifest files
+
+            This is NOT:
+            - Stream-internal offset (first packet timestamp relative to stream start)
+            - Absolute/wall-clock time
+            - Recording duration
+
+            ALTERNATIVE APPROACH (filename parsing):
+            Daily.co filenames contain Unix timestamps (milliseconds):
+            Format: {recording_start_ts}-{participant_id}-cam-audio-{track_start_ts}.webm
+            Example: 1760988935484-52f7f48b-fbab-431f-9a50-87b9abfc8255-cam-audio-1760988935922.webm
+
+            Can calculate offset: (track_start_ts - recording_start_ts) / 1000
+            - Track 0: (1760988935922 - 1760988935484) / 1000 = 0.438s
+            - Track 1: (1760988943823 - 1760988935484) / 1000 = 8.339s
+
+            TIME DIFFERENCE: PyAV metadata vs filename timestamps differ by ~209ms:
+            - Track 0: filename=438ms, metadata=229ms (diff: 209ms)
+            - Track 1: filename=8339ms, metadata=8130ms (diff: 209ms)
+
+            Consistent delta suggests network/encoding delay. PyAV metadata is ground truth
+            (represents when audio stream actually started vs when file upload initiated).
+
+            Example with 2 participants:
+                Track A: start_time=0.2s → Joined 200ms after recording began
+                Track B: start_time=8.1s → Joined 8.1 seconds later
+
+                After padding:
+                    Track A: [0.2s silence] + [speech...]
+                    Track B: [8.1s silence] + [speech...]
+
+                Whisper transcription timestamps are now synchronized:
+                    Track A word at 5.0s → happened at meeting t=5.0s
+                    Track B word at 10.0s → happened at meeting t=10.0s
+
+                Merging just sorts by timestamp - no offset calculation needed.
+
+        Padding coincidentally involves re-encoding. It's important when we work with Daily.co + Whisper.
+        This is because Daily.co returns recordings with skipped frames e.g. when microphone muted.
+        Daily.co doesn't understand those frames and ignores them, causing timestamp issues in transcription.
+        Re-encoding restores those frames. We do padding and re-encoding together just because it's convenient and more performant:
+        we need padded values for mix mp3 anyways
+        """
+
+        transcript = await self.get_transcript()
+
+        try:
+            # PyAV streams input from S3 URL efficiently (2-5MB fixed overhead for codec/filters)
+            with av.open(track_url) as in_container:
+                start_time_seconds = self._extract_stream_start_time_from_container(
+                    in_container, track_idx
+                )
+
+                if start_time_seconds <= 0:
+                    self.logger.info(
+                        f"Track {track_idx} requires no padding (start_time={start_time_seconds}s)",
+                        track_idx=track_idx,
+                    )
+                    return track_url
+
+                # Use tempfile instead of BytesIO for better memory efficiency
+                # Reduces peak memory usage during encoding/upload
+                with tempfile.NamedTemporaryFile(
+                    suffix=".webm", delete=False
+                ) as temp_file:
+                    temp_path = temp_file.name
+
+                try:
+                    self._apply_audio_padding_to_file(
+                        in_container, temp_path, start_time_seconds, track_idx
+                    )
+
+                    storage_path = (
+                        f"file_pipeline/{transcript.id}/tracks/padded_{track_idx}.webm"
+                    )
+
+                    # Upload using file handle for streaming
+                    with open(temp_path, "rb") as padded_file:
+                        await storage.put_file(storage_path, padded_file)
+                finally:
+                    # Clean up temp file
+                    Path(temp_path).unlink(missing_ok=True)
+
+                padded_url = await storage.get_file_url(
+                    storage_path,
+                    operation="get_object",
+                    expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+                )
+
+                self.logger.info(
+                    f"Successfully padded track {track_idx}",
+                    track_idx=track_idx,
+                    start_time_seconds=start_time_seconds,
+                    padded_url=padded_url,
+                )
+
+                return padded_url
+
+        except Exception as e:
+            self.logger.error(
+                f"Failed to process track {track_idx}",
+                track_idx=track_idx,
+                url=track_url,
+                error=str(e),
+                exc_info=True,
+            )
+            raise Exception(
+                f"Track {track_idx} padding failed - transcript would have incorrect timestamps"
+            ) from e
+
+    def _extract_stream_start_time_from_container(
+        self, container, track_idx: int
+    ) -> float:
+        """
+        Extract meeting-relative start time from WebM stream metadata.
+        Uses PyAV to read stream.start_time from WebM container.
+        More accurate than filename timestamps by ~209ms due to network/encoding delays.
+        """
+        start_time_seconds = 0.0
+        try:
+            audio_streams = [s for s in container.streams if s.type == "audio"]
+            stream = audio_streams[0] if audio_streams else container.streams[0]
+
+            # 1) Try stream-level start_time (most reliable for Daily.co tracks)
+            if stream.start_time is not None and stream.time_base is not None:
+                start_time_seconds = float(stream.start_time * stream.time_base)
+
+            # 2) Fallback to container-level start_time (in av.time_base units)
+            if (start_time_seconds <= 0) and (container.start_time is not None):
+                start_time_seconds = float(container.start_time * av.time_base)
+
+            # 3) Fallback to first packet DTS in stream.time_base
+            if start_time_seconds <= 0:
+                for packet in container.demux(stream):
+                    if packet.dts is not None:
+                        start_time_seconds = float(packet.dts * stream.time_base)
+                        break
+        except Exception as e:
+            self.logger.warning(
+                "PyAV metadata read failed; assuming 0 start_time",
+                track_idx=track_idx,
+                error=str(e),
+            )
+            start_time_seconds = 0.0
+
+        self.logger.info(
+            f"Track {track_idx} stream metadata: start_time={start_time_seconds:.3f}s",
+            track_idx=track_idx,
+        )
+        return start_time_seconds
+
+    def _apply_audio_padding_to_file(
+        self,
+        in_container,
+        output_path: str,
+        start_time_seconds: float,
+        track_idx: int,
+    ) -> None:
+        """Apply silence padding to audio track using PyAV filter graph, writing to file"""
+        delay_ms = math.floor(start_time_seconds * 1000)
+
+        self.logger.info(
+            f"Padding track {track_idx} with {delay_ms}ms delay using PyAV",
+            track_idx=track_idx,
+            delay_ms=delay_ms,
+        )
+
+        try:
+            with av.open(output_path, "w", format="webm") as out_container:
+                in_stream = next(
+                    (s for s in in_container.streams if s.type == "audio"), None
+                )
+                if in_stream is None:
+                    raise Exception("No audio stream in input")
+
+                out_stream = out_container.add_stream(
+                    "libopus", rate=OPUS_STANDARD_SAMPLE_RATE
+                )
+                out_stream.bit_rate = OPUS_DEFAULT_BIT_RATE
+                graph = av.filter.Graph()
+
+                abuf_args = (
+                    f"time_base=1/{OPUS_STANDARD_SAMPLE_RATE}:"
+                    f"sample_rate={OPUS_STANDARD_SAMPLE_RATE}:"
+                    f"sample_fmt=s16:"
+                    f"channel_layout=stereo"
+                )
+                src = graph.add("abuffer", args=abuf_args, name="src")
+                aresample_f = graph.add("aresample", args="async=1", name="ares")
+                # adelay requires one delay value per channel separated by '|'
+                delays_arg = f"{delay_ms}|{delay_ms}"
+                adelay_f = graph.add(
+                    "adelay", args=f"delays={delays_arg}:all=1", name="delay"
+                )
+                sink = graph.add("abuffersink", name="sink")
+
+                src.link_to(aresample_f)
+                aresample_f.link_to(adelay_f)
+                adelay_f.link_to(sink)
+                graph.configure()
+
+                resampler = AudioResampler(
+                    format="s16", layout="stereo", rate=OPUS_STANDARD_SAMPLE_RATE
+                )
+                # Decode -> resample -> push through graph -> encode Opus
+                for frame in in_container.decode(in_stream):
+                    out_frames = resampler.resample(frame) or []
+                    for rframe in out_frames:
+                        rframe.sample_rate = OPUS_STANDARD_SAMPLE_RATE
+                        rframe.time_base = Fraction(1, OPUS_STANDARD_SAMPLE_RATE)
+                        src.push(rframe)
+
+                        while True:
+                            try:
+                                f_out = sink.pull()
+                            except Exception:
+                                break
+                            f_out.sample_rate = OPUS_STANDARD_SAMPLE_RATE
+                            f_out.time_base = Fraction(1, OPUS_STANDARD_SAMPLE_RATE)
+                            for packet in out_stream.encode(f_out):
+                                out_container.mux(packet)
+
+                src.push(None)
+                while True:
+                    try:
+                        f_out = sink.pull()
+                    except Exception:
+                        break
+                    f_out.sample_rate = OPUS_STANDARD_SAMPLE_RATE
+                    f_out.time_base = Fraction(1, OPUS_STANDARD_SAMPLE_RATE)
+                    for packet in out_stream.encode(f_out):
+                        out_container.mux(packet)
+
+                for packet in out_stream.encode(None):
+                    out_container.mux(packet)
+        except Exception as e:
+            self.logger.error(
+                "PyAV padding failed for track",
+                track_idx=track_idx,
+                delay_ms=delay_ms,
+                error=str(e),
+                exc_info=True,
+            )
+            raise
+
+    async def mixdown_tracks(
+        self,
+        track_urls: list[str],
+        writer: AudioFileWriterProcessor,
+        offsets_seconds: list[float] | None = None,
+    ) -> None:
+        """Multi-track mixdown using PyAV filter graph (amix), reading from S3 presigned URLs"""
+
+        target_sample_rate: int | None = None
+        for url in track_urls:
+            if not url:
+                continue
+            container = None
+            try:
+                container = av.open(url)
+                for frame in container.decode(audio=0):
+                    target_sample_rate = frame.sample_rate
+                    break
+            except Exception:
+                continue
+            finally:
+                if container is not None:
+                    container.close()
+            if target_sample_rate:
+                break
+
+        if not target_sample_rate:
+            self.logger.error("Mixdown failed - no decodable audio frames found")
+            raise Exception("Mixdown failed: No decodable audio frames in any track")
+        # Build PyAV filter graph:
+        # N abuffer (s32/stereo)
+        #   -> optional adelay per input (for alignment)
+        #   -> amix (s32)
+        #   -> aformat(s16)
+        #   -> sink
+        graph = av.filter.Graph()
+        inputs = []
+        valid_track_urls = [url for url in track_urls if url]
+        input_offsets_seconds = None
+        if offsets_seconds is not None:
+            input_offsets_seconds = [
+                offsets_seconds[i] for i, url in enumerate(track_urls) if url
+            ]
+        for idx, url in enumerate(valid_track_urls):
+            args = (
+                f"time_base=1/{target_sample_rate}:"
+                f"sample_rate={target_sample_rate}:"
+                f"sample_fmt=s32:"
+                f"channel_layout=stereo"
+            )
+            in_ctx = graph.add("abuffer", args=args, name=f"in{idx}")
+            inputs.append(in_ctx)
+
+        if not inputs:
+            self.logger.error("Mixdown failed - no valid inputs for graph")
+            raise Exception("Mixdown failed: No valid inputs for filter graph")
+
+        mixer = graph.add("amix", args=f"inputs={len(inputs)}:normalize=0", name="mix")
+
+        fmt = graph.add(
+            "aformat",
+            args=(
+                f"sample_fmts=s32:channel_layouts=stereo:sample_rates={target_sample_rate}"
+            ),
+            name="fmt",
+        )
+
+        sink = graph.add("abuffersink", name="out")
+
+        # Optional per-input delay before mixing
+        delays_ms: list[int] = []
+        if input_offsets_seconds is not None:
+            base = min(input_offsets_seconds) if input_offsets_seconds else 0.0
+            delays_ms = [
+                max(0, int(round((o - base) * 1000))) for o in input_offsets_seconds
+            ]
+        else:
+            delays_ms = [0 for _ in inputs]
+
+        for idx, in_ctx in enumerate(inputs):
+            delay_ms = delays_ms[idx] if idx < len(delays_ms) else 0
+            if delay_ms > 0:
+                # adelay requires one value per channel; use same for stereo
+                adelay = graph.add(
+                    "adelay",
+                    args=f"delays={delay_ms}|{delay_ms}:all=1",
+                    name=f"delay{idx}",
+                )
+                in_ctx.link_to(adelay)
+                adelay.link_to(mixer, 0, idx)
+            else:
+                in_ctx.link_to(mixer, 0, idx)
+        mixer.link_to(fmt)
+        fmt.link_to(sink)
+        graph.configure()
+
+        containers = []
+        try:
+            # Open all containers with cleanup guaranteed
+            for i, url in enumerate(valid_track_urls):
+                try:
+                    c = av.open(url)
+                    containers.append(c)
+                except Exception as e:
+                    self.logger.warning(
+                        "Mixdown: failed to open container from URL",
+                        input=i,
+                        url=url,
+                        error=str(e),
+                    )
+
+            if not containers:
+                self.logger.error("Mixdown failed - no valid containers opened")
+                raise Exception("Mixdown failed: Could not open any track containers")
+
+            decoders = [c.decode(audio=0) for c in containers]
+            active = [True] * len(decoders)
+            resamplers = [
+                AudioResampler(format="s32", layout="stereo", rate=target_sample_rate)
+                for _ in decoders
+            ]
+
+            while any(active):
+                for i, (dec, is_active) in enumerate(zip(decoders, active)):
+                    if not is_active:
+                        continue
+                    try:
+                        frame = next(dec)
+                    except StopIteration:
+                        active[i] = False
+                        continue
+
+                    if frame.sample_rate != target_sample_rate:
+                        continue
+                    out_frames = resamplers[i].resample(frame) or []
+                    for rf in out_frames:
+                        rf.sample_rate = target_sample_rate
+                        rf.time_base = Fraction(1, target_sample_rate)
+                        inputs[i].push(rf)
+
+                    while True:
+                        try:
+                            mixed = sink.pull()
+                        except Exception:
+                            break
+                        mixed.sample_rate = target_sample_rate
+                        mixed.time_base = Fraction(1, target_sample_rate)
+                        await writer.push(mixed)
+
+            for in_ctx in inputs:
+                in_ctx.push(None)
+            while True:
+                try:
+                    mixed = sink.pull()
+                except Exception:
+                    break
+                mixed.sample_rate = target_sample_rate
+                mixed.time_base = Fraction(1, target_sample_rate)
+                await writer.push(mixed)
+        finally:
+            # Cleanup all containers, even if processing failed
+            for c in containers:
+                if c is not None:
+                    try:
+                        c.close()
+                    except Exception:
+                        pass  # Best effort cleanup
+
+    @broadcast_to_sockets
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            return await transcripts_controller.set_status(transcript_id, status)
+
+    async def on_waveform(self, data):
+        async with self.transaction():
+            waveform = TranscriptWaveform(waveform=data)
+            transcript = await self.get_transcript()
+            return await transcripts_controller.append_event(
+                transcript=transcript, event="WAVEFORM", data=waveform
+            )
+
+    async def update_participants_from_daily(
+        self, transcript: Transcript, track_keys: list[str]
+    ) -> None:
+        """Update transcript participants with user_id and names from Daily.co API."""
+        if not transcript.recording_id:
+            return
+
+        try:
+            async with create_platform_client("daily") as daily_client:
+                id_to_name = {}
+                id_to_user_id = {}
+
+                try:
+                    rec_details = await daily_client.get_recording(
+                        transcript.recording_id
+                    )
+                    mtg_session_id = rec_details.mtgSessionId
+                    if mtg_session_id:
+                        try:
+                            payload: MeetingParticipantsResponse = (
+                                await daily_client.get_meeting_participants(
+                                    mtg_session_id
+                                )
+                            )
+                            for p in payload.data:
+                                pid = p.participant_id
+                                name = p.user_name
+                                user_id = p.user_id
+                                if name:
+                                    id_to_name[pid] = name
+                                if user_id:
+                                    id_to_user_id[pid] = user_id
+                        except Exception as e:
+                            self.logger.warning(
+                                "Failed to fetch Daily meeting participants",
+                                error=str(e),
+                                mtg_session_id=mtg_session_id,
+                                exc_info=True,
+                            )
+                    else:
+                        self.logger.warning(
+                            "No mtgSessionId found for recording; participant names may be generic",
+                            recording_id=transcript.recording_id,
+                        )
+                except Exception as e:
+                    self.logger.warning(
+                        "Failed to fetch Daily recording details",
+                        error=str(e),
+                        recording_id=transcript.recording_id,
+                        exc_info=True,
+                    )
+                    return
+
+                cam_audio_keys = filter_cam_audio_tracks(track_keys)
+
+                for idx, key in enumerate(cam_audio_keys):
+                    try:
+                        parsed = parse_daily_recording_filename(key)
+                        participant_id = parsed.participant_id
+                    except ValueError as e:
+                        self.logger.error(
+                            "Failed to parse Daily recording filename",
+                            error=str(e),
+                            key=key,
+                            exc_info=True,
+                        )
+                        continue
+
+                    default_name = f"Speaker {idx}"
+                    name = id_to_name.get(participant_id, default_name)
+                    user_id = id_to_user_id.get(participant_id)
+
+                    participant = TranscriptParticipant(
+                        id=participant_id, speaker=idx, name=name, user_id=user_id
+                    )
+                    await transcripts_controller.upsert_participant(
+                        transcript, participant
+                    )
+
+        except Exception as e:
+            self.logger.warning(
+                "Failed to map participant names", error=str(e), exc_info=True
+            )
+
+    async def process(self, bucket_name: str, track_keys: list[str]):
+        transcript = await self.get_transcript()
+        async with self.transaction():
+            await transcripts_controller.update(
+                transcript,
+                {
+                    "events": [],
+                    "topics": [],
+                    "participants": [],
+                },
+            )
+
+        await self.update_participants_from_daily(transcript, track_keys)
+
+        source_storage = get_transcripts_storage()
+        transcript_storage = source_storage
+
+        track_urls: list[str] = []
+        for key in track_keys:
+            url = await source_storage.get_file_url(
+                key,
+                operation="get_object",
+                expires_in=PRESIGNED_URL_EXPIRATION_SECONDS,
+                bucket=bucket_name,
+            )
+            track_urls.append(url)
+            self.logger.info(
+                f"Generated presigned URL for track from {bucket_name}",
+                key=key,
+            )
+
+        created_padded_files = set()
+        padded_track_urls: list[str] = []
+        for idx, url in enumerate(track_urls):
+            padded_url = await self.pad_track_for_transcription(
+                url, idx, transcript_storage
+            )
+            padded_track_urls.append(padded_url)
+            if padded_url != url:
+                storage_path = f"file_pipeline/{transcript.id}/tracks/padded_{idx}.webm"
+                created_padded_files.add(storage_path)
+            self.logger.info(f"Track {idx} processed, padded URL: {padded_url}")
+
+        transcript.data_path.mkdir(parents=True, exist_ok=True)
+
+        if settings.SKIP_MIXDOWN:
+            self.logger.warning(
+                "SKIP_MIXDOWN enabled: Skipping mixdown and waveform generation. "
+                "UI will have no audio playback or waveform.",
+                num_tracks=len(padded_track_urls),
+                transcript_id=transcript.id,
+            )
+        else:
+            mp3_writer = AudioFileWriterProcessor(
+                path=str(transcript.audio_mp3_filename),
+                on_duration=self.on_duration,
+            )
+            await self.mixdown_tracks(
+                padded_track_urls, mp3_writer, offsets_seconds=None
+            )
+            await mp3_writer.flush()
+
+            if not transcript.audio_mp3_filename.exists():
+                raise Exception(
+                    "Mixdown failed - no MP3 file generated. Cannot proceed without playable audio."
+                )
+
+            storage_path = f"{transcript.id}/audio.mp3"
+            # Use file handle streaming to avoid loading entire MP3 into memory
+            mp3_size = transcript.audio_mp3_filename.stat().st_size
+            with open(transcript.audio_mp3_filename, "rb") as mp3_file:
+                await transcript_storage.put_file(storage_path, mp3_file)
+            mp3_url = await transcript_storage.get_file_url(storage_path)
+
+            await transcripts_controller.update(
+                transcript, {"audio_location": "storage"}
+            )
+
+            self.logger.info(
+                f"Uploaded mixed audio to storage",
+                storage_path=storage_path,
+                size=mp3_size,
+                url=mp3_url,
+            )
+
+            self.logger.info("Generating waveform from mixed audio")
+            waveform_processor = AudioWaveformProcessor(
+                audio_path=transcript.audio_mp3_filename,
+                waveform_path=transcript.audio_waveform_filename,
+                on_waveform=self.on_waveform,
+            )
+            waveform_processor.set_pipeline(self.empty_pipeline)
+            await waveform_processor.flush()
+            self.logger.info("Waveform generated successfully")
+
+        speaker_transcripts: list[TranscriptType] = []
+        for idx, padded_url in enumerate(padded_track_urls):
+            if not padded_url:
+                continue
+
+            t = await self.transcribe_file(padded_url, transcript.source_language)
+
+            if not t.words:
+                self.logger.debug(f"no words in track {idx}")
+                # not skipping, it may be silence or indistinguishable mumbling
+
+            for w in t.words:
+                w.speaker = idx
+
+            speaker_transcripts.append(t)
+            self.logger.info(
+                f"Track {idx} transcribed successfully with {len(t.words)} words",
+                track_idx=idx,
+            )
+
+        valid_track_count = len([url for url in padded_track_urls if url])
+        if valid_track_count > 0 and len(speaker_transcripts) != valid_track_count:
+            raise Exception(
+                f"Only {len(speaker_transcripts)}/{valid_track_count} tracks transcribed successfully. "
+                f"All tracks must succeed to avoid incomplete transcripts."
+            )
+
+        if not speaker_transcripts:
+            raise Exception("No valid track transcriptions")
+
+        self.logger.info(f"Cleaning up {len(created_padded_files)} temporary S3 files")
+        cleanup_tasks = []
+        for storage_path in created_padded_files:
+            cleanup_tasks.append(transcript_storage.delete_file(storage_path))
+
+        if cleanup_tasks:
+            cleanup_results = await asyncio.gather(
+                *cleanup_tasks, return_exceptions=True
+            )
+            for storage_path, result in zip(created_padded_files, cleanup_results):
+                if isinstance(result, Exception):
+                    self.logger.warning(
+                        "Failed to cleanup temporary padded track",
+                        storage_path=storage_path,
+                        error=str(result),
+                    )
+
+        merged_words = []
+        for t in speaker_transcripts:
+            merged_words.extend(t.words)
+        merged_words.sort(
+            key=lambda w: w.start if hasattr(w, "start") and w.start is not None else 0
+        )
+
+        merged_transcript = TranscriptType(words=merged_words, translation=None)
+
+        await self.on_transcript(merged_transcript)
+
+        topics = await self.detect_topics(merged_transcript, transcript.target_language)
+        await asyncio.gather(
+            self.generate_title(topics),
+            self.generate_summaries(topics),
+            return_exceptions=False,
+        )
+
+        await self.set_status(transcript.id, "ended")
+
+    async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
+        return await transcribe_file_with_processor(audio_url, language)
+
+    async def detect_topics(
+        self, transcript: TranscriptType, target_language: str
+    ) -> list[TitleSummary]:
+        return await topic_processing.detect_topics(
+            transcript,
+            target_language,
+            on_topic_callback=self.on_topic,
+            empty_pipeline=self.empty_pipeline,
+        )
+
+    async def generate_title(self, topics: list[TitleSummary]):
+        return await topic_processing.generate_title(
+            topics,
+            on_title_callback=self.on_title,
+            empty_pipeline=self.empty_pipeline,
+            logger=self.logger,
+        )
+
+    async def generate_summaries(self, topics: list[TitleSummary]):
+        transcript = await self.get_transcript()
+        return await topic_processing.generate_summaries(
+            topics,
+            transcript,
+            on_long_summary_callback=self.on_long_summary,
+            on_short_summary_callback=self.on_short_summary,
+            empty_pipeline=self.empty_pipeline,
+            logger=self.logger,
+        )
+
+
+@shared_task
+@asynctask
+async def task_pipeline_multitrack_process(
+    *, transcript_id: str, bucket_name: str, track_keys: list[str]
+):
+    pipeline = PipelineMainMultitrack(transcript_id=transcript_id)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+        await pipeline.process(bucket_name, track_keys)
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/topic_processing.py

@@ -0,0 +1,109 @@
+"""
+Topic processing utilities
+==========================
+
+Shared topic detection, title generation, and summarization logic
+used across file and multitrack pipelines.
+"""
+
+from typing import Callable
+
+import structlog
+
+from reflector.db.transcripts import Transcript
+from reflector.processors import (
+    TranscriptFinalSummaryProcessor,
+    TranscriptFinalTitleProcessor,
+    TranscriptTopicDetectorProcessor,
+)
+from reflector.processors.types import TitleSummary
+from reflector.processors.types import Transcript as TranscriptType
+
+
+class EmptyPipeline:
+    def __init__(self, logger: structlog.BoundLogger):
+        self.logger = logger
+
+    def get_pref(self, k, d=None):
+        return d
+
+    async def emit(self, event):
+        pass
+
+
+async def detect_topics(
+    transcript: TranscriptType,
+    target_language: str,
+    *,
+    on_topic_callback: Callable,
+    empty_pipeline: EmptyPipeline,
+) -> list[TitleSummary]:
+    chunk_size = 300
+    topics: list[TitleSummary] = []
+
+    async def on_topic(topic: TitleSummary):
+        topics.append(topic)
+        return await on_topic_callback(topic)
+
+    topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
+    topic_detector.set_pipeline(empty_pipeline)
+
+    for i in range(0, len(transcript.words), chunk_size):
+        chunk_words = transcript.words[i : i + chunk_size]
+        if not chunk_words:
+            continue
+
+        chunk_transcript = TranscriptType(
+            words=chunk_words, translation=transcript.translation
+        )
+
+        await topic_detector.push(chunk_transcript)
+
+    await topic_detector.flush()
+    return topics
+
+
+async def generate_title(
+    topics: list[TitleSummary],
+    *,
+    on_title_callback: Callable,
+    empty_pipeline: EmptyPipeline,
+    logger: structlog.BoundLogger,
+):
+    if not topics:
+        logger.warning("No topics for title generation")
+        return
+
+    processor = TranscriptFinalTitleProcessor(callback=on_title_callback)
+    processor.set_pipeline(empty_pipeline)
+
+    for topic in topics:
+        await processor.push(topic)
+
+    await processor.flush()
+
+
+async def generate_summaries(
+    topics: list[TitleSummary],
+    transcript: Transcript,
+    *,
+    on_long_summary_callback: Callable,
+    on_short_summary_callback: Callable,
+    empty_pipeline: EmptyPipeline,
+    logger: structlog.BoundLogger,
+):
+    if not topics:
+        logger.warning("No topics for summary generation")
+        return
+
+    processor = TranscriptFinalSummaryProcessor(
+        transcript=transcript,
+        callback=on_long_summary_callback,
+        on_short_summary=on_short_summary_callback,
+    )
+    processor.set_pipeline(empty_pipeline)
+
+    for topic in topics:
+        await processor.push(topic)
+
+    await processor.flush()

server/reflector/pipelines/transcription_helpers.py

@@ -0,0 +1,34 @@
+from reflector.processors.file_transcript import FileTranscriptInput
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.types import Transcript as TranscriptType
+
+
+async def transcribe_file_with_processor(
+    audio_url: str,
+    language: str,
+    processor_name: str | None = None,
+) -> TranscriptType:
+    processor = (
+        FileTranscriptAutoProcessor(name=processor_name)
+        if processor_name
+        else FileTranscriptAutoProcessor()
+    )
+    input_data = FileTranscriptInput(audio_url=audio_url, language=language)
+
+    result: TranscriptType | None = None
+
+    async def capture_result(transcript):
+        nonlocal result
+        result = transcript
+
+    processor.on(capture_result)
+    await processor.push(input_data)
+    await processor.flush()
+
+    if not result:
+        processor_label = processor_name or "default"
+        raise ValueError(
+            f"No transcript captured from {processor_label} processor for audio: {audio_url}"
+        )
+
+    return result

server/reflector/processors/__init__.py

@@ -1,5 +1,7 @@
 from .audio_chunker import AudioChunkerProcessor  # noqa: F401
+from .audio_chunker_auto import AudioChunkerAutoProcessor  # noqa: F401
 from .audio_diarization_auto import AudioDiarizationAutoProcessor  # noqa: F401
+from .audio_downscale import AudioDownscaleProcessor  # noqa: F401
 from .audio_file_writer import AudioFileWriterProcessor  # noqa: F401
 from .audio_merge import AudioMergeProcessor  # noqa: F401
 from .audio_transcript import AudioTranscriptProcessor  # noqa: F401

server/reflector/processors/audio_chunker.py

@@ -1,340 +1,78 @@
 from typing import Optional
 
 import av
-import numpy as np
-import torch
-from silero_vad import VADIterator, load_silero_vad
+from prometheus_client import Counter, Histogram
 
 from reflector.processors.base import Processor
 
 
 class AudioChunkerProcessor(Processor):
     """
-    Assemble audio frames into chunks with VAD-based speech detection
+    Base class for assembling audio frames into chunks
     """
 
     INPUT_TYPE = av.AudioFrame
     OUTPUT_TYPE = list[av.AudioFrame]
 
-    def __init__(
-        self,
-        block_frames=256,
-        max_frames=1024,
-        vad_threshold=0.5,
-        use_onnx=False,
-        min_frames=2,
-    ):
-        super().__init__()
+    m_chunk = Histogram(
+        "audio_chunker",
+        "Time spent in AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_call = Counter(
+        "audio_chunker_call",
+        "Number of calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_success = Counter(
+        "audio_chunker_success",
+        "Number of successful calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_failure = Counter(
+        "audio_chunker_failure",
+        "Number of failed calls to AudioChunker.chunk",
+        ["backend"],
+    )
+
+    def __init__(self, *args, **kwargs):
+        name = self.__class__.__name__
+        self.m_chunk = self.m_chunk.labels(name)
+        self.m_chunk_call = self.m_chunk_call.labels(name)
+        self.m_chunk_success = self.m_chunk_success.labels(name)
+        self.m_chunk_failure = self.m_chunk_failure.labels(name)
+        super().__init__(*args, **kwargs)
         self.frames: list[av.AudioFrame] = []
-        self.block_frames = block_frames
-        self.max_frames = max_frames
-        self.vad_threshold = vad_threshold
-        self.min_frames = min_frames
-
-        # Initialize Silero VAD
-        self._init_vad(use_onnx)
-
-    def _init_vad(self, use_onnx=False):
-        """Initialize Silero VAD model"""
-        try:
-            torch.set_num_threads(1)
-            self.vad_model = load_silero_vad(onnx=use_onnx)
-            self.vad_iterator = VADIterator(self.vad_model, sampling_rate=16000)
-            self.logger.info("Silero VAD initialized successfully")
-
-        except Exception as e:
-            self.logger.error(f"Failed to initialize Silero VAD: {e}")
-            self.vad_model = None
-            self.vad_iterator = None
 
     async def _push(self, data: av.AudioFrame):
-        self.frames.append(data)
-        # print("timestamp", data.pts * data.time_base * 1000)
-
-        # Check for speech segments every 32 frames (~1 second)
-        if len(self.frames) >= 32 and len(self.frames) % 32 == 0:
-            await self._process_block()
-
-        # Safety fallback - emit if we hit max frames
-        elif len(self.frames) >= self.max_frames:
-            self.logger.warning(
-                f"AudioChunkerProcessor: Reached max frames ({self.max_frames}), "
-                f"emitting first {self.max_frames // 2} frames"
-            )
-            frames_to_emit = self.frames[: self.max_frames // 2]
-            self.frames = self.frames[self.max_frames // 2 :]
-            if len(frames_to_emit) >= self.min_frames:
-                await self.emit(frames_to_emit)
-            else:
-                self.logger.debug(
-                    f"Ignoring fallback segment with {len(frames_to_emit)} frames "
-                    f"(< {self.min_frames} minimum)"
+        """Process incoming audio frame"""
+        # Validate audio format on first frame
+        if len(self.frames) == 0:
+            if data.sample_rate != 16000 or len(data.layout.channels) != 1:
+                raise ValueError(
+                    f"AudioChunkerProcessor expects 16kHz mono audio, got {data.sample_rate}Hz "
+                    f"with {len(data.layout.channels)} channel(s). "
+                    f"Use AudioDownscaleProcessor before this processor."
                 )
 
-    async def _process_block(self):
-        # Need at least 32 frames for VAD detection (~1 second)
-        if len(self.frames) < 32 or self.vad_iterator is None:
-            return
-
-        # Processing block with current buffer size
-        # print(f"Processing block: {len(self.frames)} frames in buffer")
-
         try:
-            # Convert frames to numpy array for VAD
-            audio_array = self._frames_to_numpy(self.frames)
+            self.m_chunk_call.inc()
+            with self.m_chunk.time():
+                result = await self._chunk(data)
+            self.m_chunk_success.inc()
+            if result:
+                await self.emit(result)
+        except Exception:
+            self.m_chunk_failure.inc()
+            raise
 
-            if audio_array is None:
-                # Fallback: emit all frames if conversion failed
-                frames_to_emit = self.frames[:]
-                self.frames = []
-                if len(frames_to_emit) >= self.min_frames:
-                    await self.emit(frames_to_emit)
-                else:
-                    self.logger.debug(
-                        f"Ignoring conversion-failed segment with {len(frames_to_emit)} frames "
-                        f"(< {self.min_frames} minimum)"
-                    )
-                return
-
-            # Find complete speech segments in the buffer
-            speech_end_frame = self._find_speech_segment_end(audio_array)
-
-            if speech_end_frame is None or speech_end_frame <= 0:
-                # No speech found but buffer is getting large
-                if len(self.frames) > 512:
-                    # Check if it's all silence and can be discarded
-                    # No speech segment found, buffer at {len(self.frames)} frames
-
-                    # Could emit silence or discard old frames here
-                    # For now, keep first 256 frames and discard older silence
-                    if len(self.frames) > 768:
-                        self.logger.debug(
-                            f"Discarding {len(self.frames) - 256} old frames (likely silence)"
-                        )
-                        self.frames = self.frames[-256:]
-                return
-
-            # Calculate segment timing information
-            frames_to_emit = self.frames[:speech_end_frame]
-
-            # Get timing from av.AudioFrame
-            if frames_to_emit:
-                first_frame = frames_to_emit[0]
-                last_frame = frames_to_emit[-1]
-                sample_rate = first_frame.sample_rate
-
-                # Calculate duration
-                total_samples = sum(f.samples for f in frames_to_emit)
-                duration_seconds = total_samples / sample_rate if sample_rate > 0 else 0
-
-                # Get timestamps if available
-                start_time = (
-                    first_frame.pts * first_frame.time_base if first_frame.pts else 0
-                )
-                end_time = (
-                    last_frame.pts * last_frame.time_base if last_frame.pts else 0
-                )
-
-                # Convert to HH:MM:SS format for logging
-                def format_time(seconds):
-                    if not seconds:
-                        return "00:00:00"
-                    total_seconds = int(float(seconds))
-                    hours = total_seconds // 3600
-                    minutes = (total_seconds % 3600) // 60
-                    secs = total_seconds % 60
-                    return f"{hours:02d}:{minutes:02d}:{secs:02d}"
-
-                start_formatted = format_time(start_time)
-                end_formatted = format_time(end_time)
-
-                # Keep remaining frames for next processing
-                remaining_after = len(self.frames) - speech_end_frame
-
-                # Single structured log line
-                self.logger.info(
-                    "Speech segment found",
-                    start=start_formatted,
-                    end=end_formatted,
-                    frames=speech_end_frame,
-                    duration=round(duration_seconds, 2),
-                    buffer_before=len(self.frames),
-                    remaining=remaining_after,
-                )
-
-            # Keep remaining frames for next processing
-            self.frames = self.frames[speech_end_frame:]
-
-            # Filter out segments with too few frames
-            if len(frames_to_emit) >= self.min_frames:
-                await self.emit(frames_to_emit)
-            else:
-                self.logger.debug(
-                    f"Ignoring segment with {len(frames_to_emit)} frames "
-                    f"(< {self.min_frames} minimum)"
-                )
-
-        except Exception as e:
-            self.logger.error(f"Error in VAD processing: {e}")
-            # Fallback to simple chunking
-            if len(self.frames) >= self.block_frames:
-                frames_to_emit = self.frames[: self.block_frames]
-                self.frames = self.frames[self.block_frames :]
-                if len(frames_to_emit) >= self.min_frames:
-                    await self.emit(frames_to_emit)
-                else:
-                    self.logger.debug(
-                        f"Ignoring exception-fallback segment with {len(frames_to_emit)} frames "
-                        f"(< {self.min_frames} minimum)"
-                    )
-
-    def _frames_to_numpy(self, frames: list[av.AudioFrame]) -> Optional[np.ndarray]:
-        """Convert av.AudioFrame list to numpy array for VAD processing"""
-        if not frames:
-            return None
-
-        try:
-            first_frame = frames[0]
-            original_sample_rate = first_frame.sample_rate
-
-            audio_data = []
-            for frame in frames:
-                frame_array = frame.to_ndarray()
-
-                # Handle stereo -> mono conversion
-                if len(frame_array.shape) == 2 and frame_array.shape[0] > 1:
-                    frame_array = np.mean(frame_array, axis=0)
-                elif len(frame_array.shape) == 2:
-                    frame_array = frame_array.flatten()
-
-                audio_data.append(frame_array)
-
-            if not audio_data:
-                return None
-
-            combined_audio = np.concatenate(audio_data)
-
-            # Resample from 48kHz to 16kHz if needed
-            if original_sample_rate != 16000:
-                combined_audio = self._resample_audio(
-                    combined_audio, original_sample_rate, 16000
-                )
-
-            # Ensure float32 format
-            if combined_audio.dtype == np.int16:
-                # Normalize int16 audio to float32 in range [-1.0, 1.0]
-                combined_audio = combined_audio.astype(np.float32) / 32768.0
-            elif combined_audio.dtype != np.float32:
-                combined_audio = combined_audio.astype(np.float32)
-
-            return combined_audio
-
-        except Exception as e:
-            self.logger.error(f"Error converting frames to numpy: {e}")
-
-        return None
-
-    def _resample_audio(
-        self, audio: np.ndarray, from_sr: int, to_sr: int
-    ) -> np.ndarray:
-        """Simple linear resampling from from_sr to to_sr"""
-        if from_sr == to_sr:
-            return audio
-
-        try:
-            # Simple linear interpolation resampling
-            ratio = to_sr / from_sr
-            new_length = int(len(audio) * ratio)
-
-            # Create indices for interpolation
-            old_indices = np.linspace(0, len(audio) - 1, new_length)
-            resampled = np.interp(old_indices, np.arange(len(audio)), audio)
-
-            return resampled.astype(np.float32)
-
-        except Exception as e:
-            self.logger.error("Resampling error", exc_info=e)
-            # Fallback: simple decimation/repetition
-            if from_sr > to_sr:
-                # Downsample by taking every nth sample
-                step = from_sr // to_sr
-                return audio[::step]
-            else:
-                # Upsample by repeating samples
-                repeat = to_sr // from_sr
-                return np.repeat(audio, repeat)
-
-    def _find_speech_segment_end(self, audio_array: np.ndarray) -> Optional[int]:
-        """Find complete speech segments and return frame index at segment end"""
-        if self.vad_iterator is None or len(audio_array) == 0:
-            return None
-
-        try:
-            # Process audio in 512-sample windows for VAD
-            window_size = 512
-            min_silence_windows = 3  # Require 3 windows of silence after speech
-
-            # Track speech state
-            in_speech = False
-            speech_start = None
-            speech_end = None
-            silence_count = 0
-
-            for i in range(0, len(audio_array), window_size):
-                chunk = audio_array[i : i + window_size]
-                if len(chunk) < window_size:
-                    chunk = np.pad(chunk, (0, window_size - len(chunk)))
-
-                # Detect if this window has speech
-                speech_dict = self.vad_iterator(chunk, return_seconds=True)
-
-                # VADIterator returns dict with 'start' and 'end' when speech segments are detected
-                if speech_dict:
-                    if not in_speech:
-                        # Speech started
-                        speech_start = i
-                        in_speech = True
-                        # Debug: print(f"Speech START at sample {i}, VAD: {speech_dict}")
-                    silence_count = 0  # Reset silence counter
-                    continue
-
-                if not in_speech:
-                    continue
-
-                # We're in speech but found silence
-                silence_count += 1
-                if silence_count < min_silence_windows:
-                    continue
-
-                # Found end of speech segment
-                speech_end = i - (min_silence_windows - 1) * window_size
-                # Debug: print(f"Speech END at sample {speech_end}")
-
-                # Convert sample position to frame index
-                samples_per_frame = self.frames[0].samples if self.frames else 1024
-                # Account for resampling: we process at 16kHz but frames might be 48kHz
-                resample_ratio = 48000 / 16000  # 3x
-                actual_sample_pos = int(speech_end * resample_ratio)
-                frame_index = actual_sample_pos // samples_per_frame
-
-                # Ensure we don't exceed buffer
-                frame_index = min(frame_index, len(self.frames))
-                return frame_index
-
-            return None
-
-        except Exception as e:
-            self.logger.error(f"Error finding speech segment: {e}")
-            return None
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """
+        Process audio frame and return chunk when ready.
+        Subclasses should implement their chunking logic here.
+        """
+        raise NotImplementedError
 
     async def _flush(self):
-        frames = self.frames[:]
-        self.frames = []
-        if frames:
-            if len(frames) >= self.min_frames:
-                await self.emit(frames)
-            else:
-                self.logger.debug(
-                    f"Ignoring flush segment with {len(frames)} frames "
-                    f"(< {self.min_frames} minimum)"
-                )
+        """Flush any remaining frames when processing ends"""
+        raise NotImplementedError

server/reflector/processors/audio_chunker_auto.py

@@ -0,0 +1,32 @@
+import importlib
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.settings import settings
+
+
+class AudioChunkerAutoProcessor(AudioChunkerProcessor):
+    _registry = {}
+
+    @classmethod
+    def register(cls, name, kclass):
+        cls._registry[name] = kclass
+
+    def __new__(cls, name: str | None = None, **kwargs):
+        if name is None:
+            name = settings.AUDIO_CHUNKER_BACKEND
+        if name not in cls._registry:
+            module_name = f"reflector.processors.audio_chunker_{name}"
+            importlib.import_module(module_name)
+
+        # gather specific configuration for the processor
+        # search `AUDIO_CHUNKER_BACKEND_XXX_YYY`, push to constructor as `backend_xxx_yyy`
+        config = {}
+        name_upper = name.upper()
+        settings_prefix = "AUDIO_CHUNKER_"
+        config_prefix = f"{settings_prefix}{name_upper}_"
+        for key, value in settings:
+            if key.startswith(config_prefix):
+                config_name = key[len(settings_prefix) :].lower()
+                config[config_name] = value
+
+        return cls._registry[name](**config | kwargs)

server/reflector/processors/audio_chunker_frames.py

@@ -0,0 +1,34 @@
+from typing import Optional
+
+import av
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerFramesProcessor(AudioChunkerProcessor):
+    """
+    Simple frame-based audio chunker that emits chunks after a fixed number of frames
+    """
+
+    def __init__(self, max_frames=256, **kwargs):
+        super().__init__(**kwargs)
+        self.max_frames = max_frames
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        self.frames.append(data)
+        if len(self.frames) >= self.max_frames:
+            frames_to_emit = self.frames[:]
+            self.frames = []
+            return frames_to_emit
+
+        return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            await self.emit(frames)
+
+
+AudioChunkerAutoProcessor.register("frames", AudioChunkerFramesProcessor)

server/reflector/processors/audio_chunker_silero.py

@@ -0,0 +1,298 @@
+from typing import Optional
+
+import av
+import numpy as np
+import torch
+from silero_vad import VADIterator, load_silero_vad
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerSileroProcessor(AudioChunkerProcessor):
+    """
+    Assemble audio frames into chunks with VAD-based speech detection using Silero VAD
+    """
+
+    def __init__(
+        self,
+        block_frames=256,
+        max_frames=1024,
+        use_onnx=True,
+        min_frames=2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.block_frames = block_frames
+        self.max_frames = max_frames
+        self.min_frames = min_frames
+
+        # Initialize Silero VAD
+        self._init_vad(use_onnx)
+
+    def _init_vad(self, use_onnx=False):
+        """Initialize Silero VAD model"""
+        try:
+            torch.set_num_threads(1)
+            self.vad_model = load_silero_vad(onnx=use_onnx)
+            self.vad_iterator = VADIterator(self.vad_model, sampling_rate=16000)
+            self.logger.info("Silero VAD initialized successfully")
+
+        except Exception as e:
+            self.logger.error(f"Failed to initialize Silero VAD: {e}")
+            self.vad_model = None
+            self.vad_iterator = None
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """Process audio frame and return chunk when ready"""
+        self.frames.append(data)
+
+        # Check for speech segments every 32 frames (~1 second)
+        if len(self.frames) >= 32 and len(self.frames) % 32 == 0:
+            return await self._process_block()
+
+        # Safety fallback - emit if we hit max frames
+        elif len(self.frames) >= self.max_frames:
+            self.logger.warning(
+                f"AudioChunkerSileroProcessor: Reached max frames ({self.max_frames}), "
+                f"emitting first {self.max_frames // 2} frames"
+            )
+            frames_to_emit = self.frames[: self.max_frames // 2]
+            self.frames = self.frames[self.max_frames // 2 :]
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring fallback segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        return None
+
+    async def _process_block(self) -> Optional[list[av.AudioFrame]]:
+        # Need at least 32 frames for VAD detection (~1 second)
+        if len(self.frames) < 32 or self.vad_iterator is None:
+            return None
+
+        # Processing block with current buffer size
+        print(f"Processing block: {len(self.frames)} frames in buffer")
+
+        try:
+            # Convert frames to numpy array for VAD
+            audio_array = self._frames_to_numpy(self.frames)
+
+            if audio_array is None:
+                # Fallback: emit all frames if conversion failed
+                frames_to_emit = self.frames[:]
+                self.frames = []
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring conversion-failed segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+                return None
+
+            # Find complete speech segments in the buffer
+            speech_end_frame = self._find_speech_segment_end(audio_array)
+
+            if speech_end_frame is None or speech_end_frame <= 0:
+                # No speech found but buffer is getting large
+                if len(self.frames) > 512:
+                    # Check if it's all silence and can be discarded
+                    # No speech segment found, buffer at {len(self.frames)} frames
+
+                    # Could emit silence or discard old frames here
+                    # For now, keep first 256 frames and discard older silence
+                    if len(self.frames) > 768:
+                        self.logger.debug(
+                            f"Discarding {len(self.frames) - 256} old frames (likely silence)"
+                        )
+                        self.frames = self.frames[-256:]
+                return None
+
+            # Calculate segment timing information
+            frames_to_emit = self.frames[:speech_end_frame]
+
+            # Get timing from av.AudioFrame
+            if frames_to_emit:
+                first_frame = frames_to_emit[0]
+                last_frame = frames_to_emit[-1]
+                sample_rate = first_frame.sample_rate
+
+                # Calculate duration
+                total_samples = sum(f.samples for f in frames_to_emit)
+                duration_seconds = total_samples / sample_rate if sample_rate > 0 else 0
+
+                # Get timestamps if available
+                start_time = (
+                    first_frame.pts * first_frame.time_base if first_frame.pts else 0
+                )
+                end_time = (
+                    last_frame.pts * last_frame.time_base if last_frame.pts else 0
+                )
+
+                # Convert to HH:MM:SS format for logging
+                def format_time(seconds):
+                    if not seconds:
+                        return "00:00:00"
+                    total_seconds = int(float(seconds))
+                    hours = total_seconds // 3600
+                    minutes = (total_seconds % 3600) // 60
+                    secs = total_seconds % 60
+                    return f"{hours:02d}:{minutes:02d}:{secs:02d}"
+
+                start_formatted = format_time(start_time)
+                end_formatted = format_time(end_time)
+
+                # Keep remaining frames for next processing
+                remaining_after = len(self.frames) - speech_end_frame
+
+                # Single structured log line
+                self.logger.info(
+                    "Speech segment found",
+                    start=start_formatted,
+                    end=end_formatted,
+                    frames=speech_end_frame,
+                    duration=round(duration_seconds, 2),
+                    buffer_before=len(self.frames),
+                    remaining=remaining_after,
+                )
+
+            # Keep remaining frames for next processing
+            self.frames = self.frames[speech_end_frame:]
+
+            # Filter out segments with too few frames
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        except Exception as e:
+            self.logger.error(f"Error in VAD processing: {e}")
+            # Fallback to simple chunking
+            if len(self.frames) >= self.block_frames:
+                frames_to_emit = self.frames[: self.block_frames]
+                self.frames = self.frames[self.block_frames :]
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring exception-fallback segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+
+        return None
+
+    def _frames_to_numpy(self, frames: list[av.AudioFrame]) -> Optional[np.ndarray]:
+        """Convert av.AudioFrame list to numpy array for VAD processing"""
+        if not frames:
+            return None
+
+        try:
+            audio_data = []
+            for frame in frames:
+                frame_array = frame.to_ndarray()
+
+                if len(frame_array.shape) == 2:
+                    frame_array = frame_array.flatten()
+
+                audio_data.append(frame_array)
+
+            if not audio_data:
+                return None
+
+            combined_audio = np.concatenate(audio_data)
+
+            # Ensure float32 format
+            if combined_audio.dtype == np.int16:
+                # Normalize int16 audio to float32 in range [-1.0, 1.0]
+                combined_audio = combined_audio.astype(np.float32) / 32768.0
+            elif combined_audio.dtype != np.float32:
+                combined_audio = combined_audio.astype(np.float32)
+
+            return combined_audio
+
+        except Exception as e:
+            self.logger.error(f"Error converting frames to numpy: {e}")
+
+        return None
+
+    def _find_speech_segment_end(self, audio_array: np.ndarray) -> Optional[int]:
+        """Find complete speech segments and return frame index at segment end"""
+        if self.vad_iterator is None or len(audio_array) == 0:
+            return None
+
+        try:
+            # Process audio in 512-sample windows for VAD
+            window_size = 512
+            min_silence_windows = 3  # Require 3 windows of silence after speech
+
+            # Track speech state
+            in_speech = False
+            speech_start = None
+            speech_end = None
+            silence_count = 0
+
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(chunk, (0, window_size - len(chunk)))
+
+                # Detect if this window has speech
+                speech_dict = self.vad_iterator(chunk, return_seconds=True)
+
+                # VADIterator returns dict with 'start' and 'end' when speech segments are detected
+                if speech_dict:
+                    if not in_speech:
+                        # Speech started
+                        speech_start = i
+                        in_speech = True
+                        # Debug: print(f"Speech START at sample {i}, VAD: {speech_dict}")
+                    silence_count = 0  # Reset silence counter
+                    continue
+
+                if not in_speech:
+                    continue
+
+                # We're in speech but found silence
+                silence_count += 1
+                if silence_count < min_silence_windows:
+                    continue
+
+                # Found end of speech segment
+                speech_end = i - (min_silence_windows - 1) * window_size
+                # Debug: print(f"Speech END at sample {speech_end}")
+
+                # Convert sample position to frame index
+                samples_per_frame = self.frames[0].samples if self.frames else 1024
+                frame_index = speech_end // samples_per_frame
+
+                # Ensure we don't exceed buffer
+                frame_index = min(frame_index, len(self.frames))
+                return frame_index
+
+            return None
+
+        except Exception as e:
+            self.logger.error(f"Error finding speech segment: {e}")
+            return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            if len(frames) >= self.min_frames:
+                await self.emit(frames)
+            else:
+                self.logger.debug(
+                    f"Ignoring flush segment with {len(frames)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+
+AudioChunkerAutoProcessor.register("silero", AudioChunkerSileroProcessor)

server/reflector/processors/audio_downscale.py

@@ -0,0 +1,60 @@
+from typing import Optional
+
+import av
+from av.audio.resampler import AudioResampler
+
+from reflector.processors.base import Processor
+
+
+def copy_frame(frame: av.AudioFrame) -> av.AudioFrame:
+    frame_copy = frame.from_ndarray(
+        frame.to_ndarray(),
+        format=frame.format.name,
+        layout=frame.layout.name,
+    )
+    frame_copy.sample_rate = frame.sample_rate
+    frame_copy.pts = frame.pts
+    frame_copy.time_base = frame.time_base
+    return frame_copy
+
+
+class AudioDownscaleProcessor(Processor):
+    """
+    Downscale audio frames to 16kHz mono format
+    """
+
+    INPUT_TYPE = av.AudioFrame
+    OUTPUT_TYPE = av.AudioFrame
+
+    def __init__(self, target_rate: int = 16000, target_layout: str = "mono", **kwargs):
+        super().__init__(**kwargs)
+        self.target_rate = target_rate
+        self.target_layout = target_layout
+        self.resampler: Optional[AudioResampler] = None
+        self.needs_resampling: Optional[bool] = None
+
+    async def _push(self, data: av.AudioFrame):
+        if self.needs_resampling is None:
+            self.needs_resampling = (
+                data.sample_rate != self.target_rate
+                or data.layout.name != self.target_layout
+            )
+
+            if self.needs_resampling:
+                self.resampler = AudioResampler(
+                    format="s16", layout=self.target_layout, rate=self.target_rate
+                )
+
+        if not self.needs_resampling or not self.resampler:
+            await self.emit(data)
+            return
+
+        resampled_frames = self.resampler.resample(copy_frame(data))
+        for resampled_frame in resampled_frames:
+            await self.emit(resampled_frame)
+
+    async def _flush(self):
+        if self.needs_resampling and self.resampler:
+            final_frames = self.resampler.resample(None)
+            for frame in final_frames:
+                await self.emit(frame)

server/reflector/processors/audio_merge.py

@@ -3,24 +3,11 @@ from time import monotonic_ns
 from uuid import uuid4
 
 import av
-from av.audio.resampler import AudioResampler
 
 from reflector.processors.base import Processor
 from reflector.processors.types import AudioFile
 
 
-def copy_frame(frame: av.AudioFrame) -> av.AudioFrame:
-    frame_copy = frame.from_ndarray(
-        frame.to_ndarray(),
-        format=frame.format.name,
-        layout=frame.layout.name,
-    )
-    frame_copy.sample_rate = frame.sample_rate
-    frame_copy.pts = frame.pts
-    frame_copy.time_base = frame.time_base
-    return frame_copy
-
-
 class AudioMergeProcessor(Processor):
     """
     Merge audio frame into a single file
@@ -29,9 +16,8 @@ class AudioMergeProcessor(Processor):
     INPUT_TYPE = list[av.AudioFrame]
     OUTPUT_TYPE = AudioFile
 
-    def __init__(self, downsample_to_16k_mono: bool = True, **kwargs):
+    def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        self.downsample_to_16k_mono = downsample_to_16k_mono
 
     async def _push(self, data: list[av.AudioFrame]):
         if not data:
@@ -39,72 +25,27 @@ class AudioMergeProcessor(Processor):
 
         # get audio information from first frame
         frame = data[0]
-        original_channels = len(frame.layout.channels)
-        original_sample_rate = frame.sample_rate
-        original_sample_width = frame.format.bytes
-
-        # determine if we need processing
-        needs_processing = self.downsample_to_16k_mono and (
-            original_sample_rate != 16000 or original_channels != 1
-        )
-
-        # determine output parameters
-        if self.downsample_to_16k_mono:
-            output_sample_rate = 16000
-            output_channels = 1
-            output_sample_width = 2  # 16-bit = 2 bytes
-        else:
-            output_sample_rate = original_sample_rate
-            output_channels = original_channels
-            output_sample_width = original_sample_width
+        output_channels = len(frame.layout.channels)
+        output_sample_rate = frame.sample_rate
+        output_sample_width = frame.format.bytes
 
         # create audio file
         uu = uuid4().hex
         fd = io.BytesIO()
 
-        if needs_processing:
-            # Process with PyAV resampler
-            out_container = av.open(fd, "w", format="wav")
-            out_stream = out_container.add_stream("pcm_s16le", rate=16000)
-            out_stream.layout = "mono"
+        # Use PyAV to write frames
+        out_container = av.open(fd, "w", format="wav")
+        out_stream = out_container.add_stream("pcm_s16le", rate=output_sample_rate)
+        out_stream.layout = frame.layout.name
 
-            # Create resampler if needed
-            resampler = None
-            if original_sample_rate != 16000 or original_channels != 1:
-                resampler = AudioResampler(format="s16", layout="mono", rate=16000)
-
-            for frame in data:
-                if resampler:
-                    # Resample and convert to mono
-                    # XXX for an unknown reason, if we don't use a copy of the frame, we get
-                    # Invalid Argumment from resample. Debugging indicate that when a previous processor
-                    # already used the frame (like AudioFileWriter), it make it invalid argument here.
-                    resampled_frames = resampler.resample(copy_frame(frame))
-                    for resampled_frame in resampled_frames:
-                        for packet in out_stream.encode(resampled_frame):
-                            out_container.mux(packet)
-                else:
-                    # Direct encoding without resampling
-                    for packet in out_stream.encode(frame):
-                        out_container.mux(packet)
-
-            # Flush the encoder
-            for packet in out_stream.encode(None):
+        for frame in data:
+            for packet in out_stream.encode(frame):
                 out_container.mux(packet)
-            out_container.close()
-        else:
-            # Use PyAV for original frames (no processing needed)
-            out_container = av.open(fd, "w", format="wav")
-            out_stream = out_container.add_stream("pcm_s16le", rate=output_sample_rate)
-            out_stream.layout = "mono" if output_channels == 1 else frame.layout
 
-            for frame in data:
-                for packet in out_stream.encode(frame):
-                    out_container.mux(packet)
-
-            for packet in out_stream.encode(None):
-                out_container.mux(packet)
-            out_container.close()
+        # Flush the encoder
+        for packet in out_stream.encode(None):
+            out_container.mux(packet)
+        out_container.close()
 
         fd.seek(0)
 

server/reflector/processors/audio_transcript_modal.py

@@ -12,9 +12,6 @@ API will be a POST request to TRANSCRIPT_URL:
 
 """
 
-from typing import List
-
-import aiohttp
 from openai import AsyncOpenAI
 
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
@@ -25,7 +22,9 @@ from reflector.settings import settings
 
 class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
     def __init__(
-        self, modal_api_key: str | None = None, batch_enabled: bool = True, **kwargs
+        self,
+        modal_api_key: str | None = None,
+        **kwargs,
     ):
         super().__init__()
         if not settings.TRANSCRIPT_URL:
@@ -35,126 +34,6 @@ class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
         self.transcript_url = settings.TRANSCRIPT_URL + "/v1"
         self.timeout = settings.TRANSCRIPT_TIMEOUT
         self.modal_api_key = modal_api_key
-        self.max_batch_duration = 10.0
-        self.max_batch_files = 15
-        self.batch_enabled = batch_enabled
-        self.pending_files: List[AudioFile] = []  # Files waiting to be processed
-
-    @classmethod
-    def _calculate_duration(cls, audio_file: AudioFile) -> float:
-        """Calculate audio duration in seconds from AudioFile metadata"""
-        # Duration = total_samples / sample_rate
-        # We need to estimate total samples from the file data
-        import wave
-
-        try:
-            # Try to read as WAV file to get duration
-            audio_file.fd.seek(0)
-            with wave.open(audio_file.fd, "rb") as wav_file:
-                frames = wav_file.getnframes()
-                sample_rate = wav_file.getframerate()
-                duration = frames / sample_rate
-                return duration
-        except Exception:
-            # Fallback: estimate from file size and audio parameters
-            audio_file.fd.seek(0, 2)  # Seek to end
-            file_size = audio_file.fd.tell()
-            audio_file.fd.seek(0)  # Reset to beginning
-
-            # Estimate: file_size / (sample_rate * channels * sample_width)
-            bytes_per_second = (
-                audio_file.sample_rate
-                * audio_file.channels
-                * (audio_file.sample_width // 8)
-            )
-            estimated_duration = (
-                file_size / bytes_per_second if bytes_per_second > 0 else 0
-            )
-            return max(0, estimated_duration)
-
-    def _create_batches(self, audio_files: List[AudioFile]) -> List[List[AudioFile]]:
-        """Group audio files into batches with maximum 30s total duration"""
-        batches = []
-        current_batch = []
-        current_duration = 0.0
-
-        for audio_file in audio_files:
-            duration = self._calculate_duration(audio_file)
-
-            # If adding this file exceeds max duration, start a new batch
-            if current_duration + duration > self.max_batch_duration and current_batch:
-                batches.append(current_batch)
-                current_batch = [audio_file]
-                current_duration = duration
-            else:
-                current_batch.append(audio_file)
-                current_duration += duration
-
-        # Add the last batch if not empty
-        if current_batch:
-            batches.append(current_batch)
-
-        return batches
-
-    async def _transcript_batch(self, audio_files: List[AudioFile]) -> List[Transcript]:
-        """Transcribe a batch of audio files using the parakeet backend"""
-        if not audio_files:
-            return []
-
-        self.logger.debug(f"Batch transcribing {len(audio_files)} files")
-
-        # Prepare form data for batch request
-        data = aiohttp.FormData()
-        data.add_field("language", self.get_pref("audio:source_language", "en"))
-        data.add_field("batch", "true")
-
-        for i, audio_file in enumerate(audio_files):
-            audio_file.fd.seek(0)
-            data.add_field(
-                "files",
-                audio_file.fd,
-                filename=f"{audio_file.name}",
-                content_type="audio/wav",
-            )
-
-        # Make batch request
-        headers = {"Authorization": f"Bearer {self.modal_api_key}"}
-
-        async with aiohttp.ClientSession(
-            timeout=aiohttp.ClientTimeout(total=self.timeout)
-        ) as session:
-            async with session.post(
-                f"{self.transcript_url}/audio/transcriptions",
-                data=data,
-                headers=headers,
-            ) as response:
-                if response.status != 200:
-                    error_text = await response.text()
-                    raise Exception(
-                        f"Batch transcription failed: {response.status} {error_text}"
-                    )
-
-                result = await response.json()
-
-        # Process batch results
-        transcripts = []
-        results = result.get("results", [])
-
-        for i, (audio_file, file_result) in enumerate(zip(audio_files, results)):
-            transcript = Transcript(
-                words=[
-                    Word(
-                        text=word_info["word"],
-                        start=word_info["start"],
-                        end=word_info["end"],
-                    )
-                    for word_info in file_result.get("words", [])
-                ]
-            )
-            transcript.add_offset(audio_file.timestamp)
-            transcripts.append(transcript)
-
-        return transcripts
 
     async def _transcript(self, data: AudioFile):
         async with AsyncOpenAI(
@@ -187,96 +66,5 @@ class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
 
         return transcript
 
-    async def transcript_multiple(
-        self, audio_files: List[AudioFile]
-    ) -> List[Transcript]:
-        """Transcribe multiple audio files using batching"""
-        if len(audio_files) == 1:
-            # Single file, use existing method
-            return [await self._transcript(audio_files[0])]
-
-        # Create batches with max 30s duration each
-        batches = self._create_batches(audio_files)
-
-        self.logger.debug(
-            f"Processing {len(audio_files)} files in {len(batches)} batches"
-        )
-
-        # Process all batches concurrently
-        all_transcripts = []
-
-        for batch in batches:
-            batch_transcripts = await self._transcript_batch(batch)
-            all_transcripts.extend(batch_transcripts)
-
-        return all_transcripts
-
-    async def _push(self, data: AudioFile):
-        """Override _push to support batching"""
-        if not self.batch_enabled:
-            # Use parent implementation for single file processing
-            return await super()._push(data)
-
-        # Add file to pending batch
-        self.pending_files.append(data)
-        self.logger.debug(
-            f"Added file to batch: {data.name}, batch size: {len(self.pending_files)}"
-        )
-
-        # Calculate total duration of pending files
-        total_duration = sum(self._calculate_duration(f) for f in self.pending_files)
-
-        # Process batch if it reaches max duration or has multiple files ready for optimization
-        should_process_batch = (
-            total_duration >= self.max_batch_duration
-            or len(self.pending_files) >= self.max_batch_files
-        )
-
-        if should_process_batch:
-            await self._process_pending_batch()
-
-    async def _process_pending_batch(self):
-        """Process all pending files as batches"""
-        if not self.pending_files:
-            return
-
-        self.logger.debug(f"Processing batch of {len(self.pending_files)} files")
-
-        try:
-            # Create batches respecting duration limit
-            batches = self._create_batches(self.pending_files)
-
-            # Process each batch
-            for batch in batches:
-                self.m_transcript_call.inc()
-                try:
-                    with self.m_transcript.time():
-                        # Use batch transcription
-                        transcripts = await self._transcript_batch(batch)
-
-                    self.m_transcript_success.inc()
-
-                    # Emit each transcript
-                    for transcript in transcripts:
-                        if transcript:
-                            await self.emit(transcript)
-
-                except Exception:
-                    self.m_transcript_failure.inc()
-                    raise
-                finally:
-                    # Release audio files
-                    for audio_file in batch:
-                        audio_file.release()
-
-        finally:
-            # Clear pending files
-            self.pending_files.clear()
-
-    async def _flush(self):
-        """Process any remaining files when flushing"""
-        await self._process_pending_batch()
-        await super()._flush()
-
 
 AudioTranscriptAutoProcessor.register("modal", AudioTranscriptModalProcessor)