feat: self-hosted gpu api (#636)

* Self-hosted gpu api * Refactor self-hosted api * Rename model api tests * Use lifespan instead of startup event * Fix self hosted imports * Add newlines * Add response models * Move gpu dir to the root * Add project description * Refactor lifespan * Update env var names for model api tests * Preload diarizarion service * Refactor uploaded file paths
2025-12-20 12:19:06 +00:00 · 2025-09-17 18:52:03 +02:00
parent fa049e8d06
commit ab859d65a6
30 changed files with 4020 additions and 16 deletions
--- a/gpu/modal_deployments/.gitignore
+++ b/gpu/modal_deployments/.gitignore
@@ -0,0 +1,33 @@
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Logs
+*.log
+
+# Env and secrets
+.env
+.env.*
+*.env
+*.secret
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Modal local state (if any)
+modal_mounts/
+.modal_cache/
--- a/gpu/modal_deployments/README.md
+++ b/gpu/modal_deployments/README.md
@@ -0,0 +1,171 @@
+# Reflector GPU implementation - Transcription and LLM
+
+This repository hold an API for the GPU implementation of the Reflector API service,
+and use [Modal.com](https://modal.com)
+
+- `reflector_diarizer.py` - Diarization API
+- `reflector_transcriber.py` - Transcription API (Whisper)
+- `reflector_transcriber_parakeet.py` - Transcription API (NVIDIA Parakeet)
+- `reflector_translator.py` - Translation API
+
+## Modal.com deployment
+
+Create a modal secret, and name it `reflector-gpu`.
+It should contain an `REFLECTOR_APIKEY` environment variable with a value.
+
+The deployment is done using [Modal.com](https://modal.com) service.
+
+```
+$ modal deploy reflector_transcriber.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-web.modal.run
+
+$ modal deploy reflector_transcriber_parakeet.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-parakeet-web.modal.run
+
+$ modal deploy reflector_llm.py
+...
+└── 🔨 Created web => https://xxxx--reflector-llm-web.modal.run
+```
+
+Then in your reflector api configuration `.env`, you can set these keys:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://xxxx--reflector-transcriber-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=REFLECTOR_APIKEY
+
+DIARIZATION_BACKEND=modal
+DIARIZATION_URL=https://xxxx--reflector-diarizer-web.modal.run
+DIARIZATION_MODAL_API_KEY=REFLECTOR_APIKEY
+
+TRANSLATION_BACKEND=modal
+TRANSLATION_URL=https://xxxx--reflector-translator-web.modal.run
+TRANSLATION_MODAL_API_KEY=REFLECTOR_APIKEY
+```
+
+## API
+
+Authentication must be passed with the `Authorization` header, using the `bearer` scheme.
+
+```
+Authorization: bearer <REFLECTOR_APIKEY>
+```
+
+### LLM
+
+`POST /llm`
+
+**request**
+```
+{
+    "prompt": "xxx"
+}
+```
+
+**response**
+```
+{
+    "text": "xxx completed"
+}
+```
+
+### Transcription
+
+#### Parakeet Transcriber (`reflector_transcriber_parakeet.py`)
+
+NVIDIA Parakeet is a state-of-the-art ASR model optimized for real-time transcription with superior word-level timestamps.
+
+**GPU Configuration:**
+- **A10G GPU** - Used for `/v1/audio/transcriptions` endpoint (small files, live transcription)
+  - Higher concurrency (max_inputs=10)
+  - Optimized for multiple small audio files
+  - Supports batch processing for efficiency
+
+- **L40S GPU** - Used for `/v1/audio/transcriptions-from-url` endpoint (large files)
+  - Lower concurrency but more powerful processing
+  - Optimized for single large audio files
+  - VAD-based chunking for long-form audio
+
+##### `/v1/audio/transcriptions` - Small file transcription
+
+**request** (multipart/form-data)
+- `file` or `files[]` - audio file(s) to transcribe
+- `model` - model name (default: `nvidia/parakeet-tdt-0.6b-v2`)
+- `language` - language code (default: `en`)
+- `batch` - whether to use batch processing for multiple files (default: `true`)
+
+**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"
+}
+```
+
+For multiple files with batch=true:
+```json
+{
+    "results": [
+        {
+            "filename": "audio1.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        },
+        {
+            "filename": "audio2.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        }
+    ]
+}
+```
+
+##### `/v1/audio/transcriptions-from-url` - Large file transcription
+
+**request** (application/json)
+```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 from large file",
+    "words": [
+        {"word": "hello", "start": 0.0, "end": 0.5},
+        {"word": "world", "start": 0.5, "end": 1.0}
+    ]
+}
+```
+
+**Supported file types:** mp3, mp4, mpeg, mpga, m4a, wav, webm
+
+#### Whisper Transcriber (`reflector_transcriber.py`)
+
+`POST /transcribe`
+
+**request** (multipart/form-data)
+
+- `file` - audio file
+- `language` - language code (e.g. `en`)
+
+**response**
+```
+{
+    "text": "xxx",
+    "words": [
+        {"text": "xxx", "start": 0.0, "end": 1.0}
+    ]
+}
+```
--- a/gpu/modal_deployments/reflector_diarizer.py
+++ b/gpu/modal_deployments/reflector_diarizer.py
@@ -0,0 +1,253 @@
+"""
+Reflector GPU backend - diarizer
+===================================
+"""
+
+import os
+import uuid
+from typing import Mapping, NewType
+from urllib.parse import urlparse
+
+import modal
+
+PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
+MODEL_DIR = "/root/diarization_models"
+UPLOADS_PATH = "/uploads"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+
+DiarizerUniqFilename = NewType("DiarizerUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App(name="reflector-diarizer")
+
+# Volume for temporary file uploads
+upload_volume = modal.Volume.from_name("diarizer-uploads", create_if_missing=True)
+
+
+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[DiarizerUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    print(f"Checking audio file at: {audio_file_url}")
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    print(f"Downloading audio file from: {audio_file_url}")
+    response = requests.get(audio_file_url, allow_redirects=True)
+
+    if response.status_code != 200:
+        print(f"Download failed with status {response.status_code}: {response.text}")
+        raise HTTPException(
+            status_code=response.status_code,
+            detail=f"Failed to download audio file: {response.status_code}",
+        )
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = DiarizerUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    print(f"Writing file to: {file_path} (size: {len(response.content)} bytes)")
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    print(f"File saved as: {unique_filename}")
+    return unique_filename, audio_suffix
+
+
+def migrate_cache_llm():
+    """
+    XXX The cache for model files in Transformers v4.22.0 has been updated.
+    Migrating your old cache. This is a one-time only operation. You can
+    interrupt this and resume the migration later on by calling
+    `transformers.utils.move_cache()`.
+    """
+    from transformers.utils.hub import move_cache
+
+    print("Moving LLM cache")
+    move_cache(cache_dir=MODEL_DIR, new_cache_dir=MODEL_DIR)
+    print("LLM cache moved")
+
+
+def download_pyannote_audio():
+    from pyannote.audio import Pipeline
+
+    Pipeline.from_pretrained(
+        PYANNOTE_MODEL_NAME,
+        cache_dir=MODEL_DIR,
+        use_auth_token=os.environ["HF_TOKEN"],
+    )
+
+
+diarizer_image = (
+    modal.Image.debian_slim(python_version="3.10.8")
+    .pip_install(
+        "pyannote.audio==3.1.0",
+        "requests",
+        "onnx",
+        "torchaudio",
+        "onnxruntime-gpu",
+        "torch==2.0.0",
+        "transformers==4.34.0",
+        "sentencepiece",
+        "protobuf",
+        "numpy",
+        "huggingface_hub",
+        "hf-transfer",
+    )
+    .run_function(
+        download_pyannote_audio,
+        secrets=[modal.Secret.from_name("hf_token")],
+    )
+    .run_function(migrate_cache_llm)
+    .env(
+        {
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
+            )
+        }
+    )
+)
+
+
+@app.cls(
+    gpu="A100",
+    timeout=60 * 30,
+    image=diarizer_image,
+    volumes={UPLOADS_PATH: upload_volume},
+    enable_memory_snapshot=True,
+    experimental_options={"enable_gpu_snapshot": True},
+    secrets=[
+        modal.Secret.from_name("hf_token"),
+    ],
+)
+@modal.concurrent(max_inputs=1)
+class Diarizer:
+    @modal.enter(snap=True)
+    def enter(self):
+        import torch
+        from pyannote.audio import Pipeline
+
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        print(f"Using device: {self.device}")
+        self.diarization_pipeline = Pipeline.from_pretrained(
+            PYANNOTE_MODEL_NAME,
+            cache_dir=MODEL_DIR,
+            use_auth_token=os.environ["HF_TOKEN"],
+        )
+        self.diarization_pipeline.to(torch.device(self.device))
+
+    @modal.method()
+    def diarize(self, filename: str, timestamp: float = 0.0):
+        import torchaudio
+
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        print(f"Diarizing audio from: {file_path}")
+        waveform, sample_rate = torchaudio.load(file_path)
+        diarization = self.diarization_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:]),
+                }
+            )
+        print("Diarization complete")
+        return {"diarization": words}
+
+
+# -------------------------------------------------------------------
+# Web API
+# -------------------------------------------------------------------
+
+
+@app.function(
+    timeout=60 * 10,
+    scaledown_window=60 * 3,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={UPLOADS_PATH: upload_volume},
+    image=diarizer_image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import Depends, FastAPI, HTTPException, status
+    from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
+
+    diarizerstub = Diarizer()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    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 DiarizationResponse(BaseModel):
+        result: dict
+
+    @app.post("/diarize", dependencies=[Depends(apikey_auth)])
+    def diarize(audio_file_url: str, timestamp: float = 0.0) -> DiarizationResponse:
+        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
+
+        try:
+            func = diarizerstub.diarize.spawn(
+                filename=unique_filename, timestamp=timestamp
+            )
+            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
--- a/gpu/modal_deployments/reflector_transcriber.py
+++ b/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()
--- a/gpu/modal_deployments/reflector_transcriber_parakeet.py
+++ b/gpu/modal_deployments/reflector_transcriber_parakeet.py
@@ -0,0 +1,658 @@
+import logging
+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 = "nvidia/parakeet-tdt-0.6b-v2"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/cache"
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+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
+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: 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"]
+        ]
+
+        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: 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"]
+            ]
+
+            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,
+        ) -> Generator[TimeSegment, None, None]:
+            """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)
+
+                    yield TimeSegment(start_time, end_time)
+                    start = None
+
+            vad_iterator.reset_states()
+
+        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]
+
+                                  ↓ (max_duration=10)
+
+              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
+
+                total_duration = end_time - batch_start_time
+
+                if total_duration <= max_duration:
+                    batch_end_time = end_time
+                    continue
+
+                yield TimeSegment(batch_start_time, batch_end_time)
+                batch_start_time = start_time
+                batch_end_time = end_time
+
+            if batch_start_time is None or batch_end_time is None:
+                return
+
+            yield TimeSegment(batch_start_time, batch_end_time)
+
+        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.
+
+            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]
+
+                padded_segment = pad_audio(audio_segment, SAMPLERATE)
+
+                yield AudioSegment(start_time, end_time, padded_segment)
+
+        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: 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, segment) in enumerate(zip(results, segments_info)):
+                start_time, end_time = segment.start, segment.end
+                text = output.text.strip()
+                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),
+                    )
+                    for word_info in output.timestamp["word"]
+                ]
+
+                yield TranscriptResult(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)
+
+        all_text_parts: list[str] = []
+        all_words: list[WordTiming] = []
+
+        raw_segments = vad_segment_generator(audio_array)
+        speech_segments = batch_speech_segments(
+            raw_segments,
+            VAD_CONFIG["batch_max_duration"],
+        )
+        audio_segments = batch_segment_to_audio_segment(speech_segments, audio_array)
+
+        for batch in audio_segments:
+            audio_segment = batch.audio
+            results = transcribe_batch(self.model, [audio_segment])
+
+            for result in emit_results(
+                results,
+                [batch],
+            ):
+                if not result.text:
+                    continue
+                all_text_parts.append(result.text)
+                all_words.extend(result.words)
+
+        all_words = enforce_word_timing_constraints(all_words)
+
+        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()
--- a/gpu/modal_deployments/reflector_translator.py
+++ b/gpu/modal_deployments/reflector_translator.py
@@ -0,0 +1,428 @@
+"""
+Reflector GPU backend - transcriber
+===================================
+"""
+
+import os
+import threading
+
+from modal import App, Image, Secret, asgi_app, enter, method
+from pydantic import BaseModel
+
+# Seamless M4T
+SEAMLESSM4T_MODEL_SIZE: str = "medium"
+SEAMLESSM4T_MODEL_CARD_NAME: str = f"seamlessM4T_{SEAMLESSM4T_MODEL_SIZE}"
+SEAMLESSM4T_VOCODER_CARD_NAME: str = "vocoder_36langs"
+
+HF_SEAMLESS_M4TEPO: str = f"facebook/seamless-m4t-{SEAMLESSM4T_MODEL_SIZE}"
+HF_SEAMLESS_M4T_VOCODEREPO: str = "facebook/seamless-m4t-vocoder"
+
+SEAMLESS_GITEPO: str = "https://github.com/facebookresearch/seamless_communication.git"
+SEAMLESS_MODEL_DIR: str = "m4t"
+
+app = App(name="reflector-translator")
+
+
+def install_seamless_communication():
+    import os
+    import subprocess
+
+    initial_dir = os.getcwd()
+    subprocess.run(
+        ["ssh-keyscan", "-t", "rsa", "github.com", ">>", "~/.ssh/known_hosts"]
+    )
+    subprocess.run(["rm", "-rf", "seamless_communication"])
+    subprocess.run(["git", "clone", SEAMLESS_GITEPO, "." + "/seamless_communication"])
+    os.chdir("seamless_communication")
+    subprocess.run(["pip", "install", "-e", "."])
+    os.chdir(initial_dir)
+
+
+def download_seamlessm4t_model():
+    from huggingface_hub import snapshot_download
+
+    print("Downloading Transcriber model & tokenizer")
+    snapshot_download(HF_SEAMLESS_M4TEPO, cache_dir=SEAMLESS_MODEL_DIR)
+    print("Transcriber model & tokenizer downloaded")
+
+    print("Downloading vocoder weights")
+    snapshot_download(HF_SEAMLESS_M4T_VOCODEREPO, cache_dir=SEAMLESS_MODEL_DIR)
+    print("Vocoder weights downloaded")
+
+
+def configure_seamless_m4t():
+    import os
+
+    import yaml
+
+    CARDS_DIR: str = "./seamless_communication/src/seamless_communication/cards"
+
+    with open(f"{CARDS_DIR}/seamlessM4T_{SEAMLESSM4T_MODEL_SIZE}.yaml", "r") as file:
+        model_yaml_data = yaml.load(file, Loader=yaml.FullLoader)
+    with open(f"{CARDS_DIR}/vocoder_36langs.yaml", "r") as file:
+        vocoder_yaml_data = yaml.load(file, Loader=yaml.FullLoader)
+    with open(f"{CARDS_DIR}/unity_nllb-100.yaml", "r") as file:
+        unity_100_yaml_data = yaml.load(file, Loader=yaml.FullLoader)
+    with open(f"{CARDS_DIR}/unity_nllb-200.yaml", "r") as file:
+        unity_200_yaml_data = yaml.load(file, Loader=yaml.FullLoader)
+
+    model_dir = f"{SEAMLESS_MODEL_DIR}/models--facebook--seamless-m4t-{SEAMLESSM4T_MODEL_SIZE}/snapshots"
+    available_model_versions = os.listdir(model_dir)
+    latest_model_version = sorted(available_model_versions)[-1]
+    model_name = f"multitask_unity_{SEAMLESSM4T_MODEL_SIZE}.pt"
+    model_path = os.path.join(os.getcwd(), model_dir, latest_model_version, model_name)
+
+    vocoder_dir = (
+        f"{SEAMLESS_MODEL_DIR}/models--facebook--seamless-m4t-vocoder/snapshots"
+    )
+    available_vocoder_versions = os.listdir(vocoder_dir)
+    latest_vocoder_version = sorted(available_vocoder_versions)[-1]
+    vocoder_name = "vocoder_36langs.pt"
+    vocoder_path = os.path.join(
+        os.getcwd(), vocoder_dir, latest_vocoder_version, vocoder_name
+    )
+
+    tokenizer_name = "tokenizer.model"
+    tokenizer_path = os.path.join(
+        os.getcwd(), model_dir, latest_model_version, tokenizer_name
+    )
+
+    model_yaml_data["checkpoint"] = f"file://{model_path}"
+    vocoder_yaml_data["checkpoint"] = f"file://{vocoder_path}"
+    unity_100_yaml_data["tokenizer"] = f"file://{tokenizer_path}"
+    unity_200_yaml_data["tokenizer"] = f"file://{tokenizer_path}"
+
+    with open(f"{CARDS_DIR}/seamlessM4T_{SEAMLESSM4T_MODEL_SIZE}.yaml", "w") as file:
+        yaml.dump(model_yaml_data, file)
+    with open(f"{CARDS_DIR}/vocoder_36langs.yaml", "w") as file:
+        yaml.dump(vocoder_yaml_data, file)
+    with open(f"{CARDS_DIR}/unity_nllb-100.yaml", "w") as file:
+        yaml.dump(unity_100_yaml_data, file)
+    with open(f"{CARDS_DIR}/unity_nllb-200.yaml", "w") as file:
+        yaml.dump(unity_200_yaml_data, file)
+
+
+transcriber_image = (
+    Image.debian_slim(python_version="3.10.8")
+    .apt_install("git")
+    .apt_install("wget")
+    .apt_install("libsndfile-dev")
+    .pip_install(
+        "requests",
+        "torch",
+        "transformers==4.34.0",
+        "sentencepiece",
+        "protobuf",
+        "huggingface_hub==0.16.4",
+        "gitpython",
+        "torchaudio",
+        "fairseq2",
+        "pyyaml",
+        "hf-transfer~=0.1",
+    )
+    .run_function(install_seamless_communication)
+    .run_function(download_seamlessm4t_model)
+    .run_function(configure_seamless_m4t)
+    .env(
+        {
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
+            )
+        }
+    )
+)
+
+
+@app.cls(
+    gpu="A10G",
+    timeout=60 * 5,
+    scaledown_window=60 * 5,
+    allow_concurrent_inputs=4,
+    image=transcriber_image,
+)
+class Translator:
+    @enter()
+    def enter(self):
+        import torch
+        from seamless_communication.inference.translator import Translator
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.translator = Translator(
+            SEAMLESSM4T_MODEL_CARD_NAME,
+            SEAMLESSM4T_VOCODER_CARD_NAME,
+            torch.device(self.device),
+            dtype=torch.float32,
+        )
+
+    @method()
+    def warmup(self):
+        return {"status": "ok"}
+
+    def get_seamless_lang_code(self, lang_code: str):
+        """
+        The codes for SeamlessM4T is different from regular standards.
+        For ex, French is "fra" and not "fr".
+        """
+        # TODO: Enhance with complete list of lang codes
+        seamless_lang_code = {
+            # Afrikaans
+            "af": "afr",
+            # Amharic
+            "am": "amh",
+            # Modern Standard Arabic
+            "ar": "arb",
+            # Moroccan Arabic
+            "ary": "ary",
+            # Egyptian Arabic
+            "arz": "arz",
+            # Assamese
+            "as": "asm",
+            # North Azerbaijani
+            "az": "azj",
+            # Belarusian
+            "be": "bel",
+            # Bengali
+            "bn": "ben",
+            # Bosnian
+            "bs": "bos",
+            # Bulgarian
+            "bg": "bul",
+            # Catalan
+            "ca": "cat",
+            # Cebuano
+            "ceb": "ceb",
+            # Czech
+            "cs": "ces",
+            # Central Kurdish
+            "ku": "ckb",
+            # Mandarin Chinese
+            "cmn": "cmn_Hant",
+            # Welsh
+            "cy": "cym",
+            # Danish
+            "da": "dan",
+            # German
+            "de": "deu",
+            # Greek
+            "el": "ell",
+            # English
+            "en": "eng",
+            # Estonian
+            "et": "est",
+            # Basque
+            "eu": "eus",
+            # Finnish
+            "fi": "fin",
+            # French
+            "fr": "fra",
+            # Irish
+            "ga": "gle",
+            # West Central Oromo,
+            "gaz": "gaz",
+            # Galician
+            "gl": "glg",
+            # Gujarati
+            "gu": "guj",
+            # Hebrew
+            "he": "heb",
+            # Hindi
+            "hi": "hin",
+            # Croatian
+            "hr": "hrv",
+            # Hungarian
+            "hu": "hun",
+            # Armenian
+            "hy": "hye",
+            # Igbo
+            "ig": "ibo",
+            # Indonesian
+            "id": "ind",
+            # Icelandic
+            "is": "isl",
+            # Italian
+            "it": "ita",
+            # Javanese
+            "jv": "jav",
+            # Japanese
+            "ja": "jpn",
+            # Kannada
+            "kn": "kan",
+            # Georgian
+            "ka": "kat",
+            # Kazakh
+            "kk": "kaz",
+            # Halh Mongolian
+            "khk": "khk",
+            # Khmer
+            "km": "khm",
+            # Kyrgyz
+            "ky": "kir",
+            # Korean
+            "ko": "kor",
+            # Lao
+            "lo": "lao",
+            # Lithuanian
+            "lt": "lit",
+            # Ganda
+            "lg": "lug",
+            # Luo
+            "luo": "luo",
+            # Standard Latvian
+            "lv": "lvs",
+            # Maithili
+            "mai": "mai",
+            # Malayalam
+            "ml": "mal",
+            # Marathi
+            "mr": "mar",
+            # Macedonian
+            "mk": "mkd",
+            # Maltese
+            "mt": "mlt",
+            # Meitei
+            "mni": "mni",
+            # Burmese
+            "my": "mya",
+            # Dutch
+            "nl": "nld",
+            # Norwegian Nynorsk
+            "nn": "nno",
+            # Norwegian Bokmål
+            "nb": "nob",
+            # Nepali
+            "ne": "npi",
+            # Nyanja
+            "ny": "nya",
+            # Odia
+            "or": "ory",
+            # Punjabi
+            "pa": "pan",
+            # Southern Pashto
+            "pbt": "pbt",
+            # Western Persian
+            "pes": "pes",
+            # Polish
+            "pl": "pol",
+            # Portuguese
+            "pt": "por",
+            # Romanian
+            "ro": "ron",
+            # Russian
+            "ru": "rus",
+            # Slovak
+            "sk": "slk",
+            # Slovenian
+            "sl": "slv",
+            # Shona
+            "sn": "sna",
+            # Sindhi
+            "sd": "snd",
+            # Somali
+            "so": "som",
+            # Spanish
+            "es": "spa",
+            # Serbian
+            "sr": "srp",
+            # Swedish
+            "sv": "swe",
+            # Swahili
+            "sw": "swh",
+            # Tamil
+            "ta": "tam",
+            # Telugu
+            "te": "tel",
+            # Tajik
+            "tg": "tgk",
+            # Tagalog
+            "tl": "tgl",
+            # Thai
+            "th": "tha",
+            # Turkish
+            "tr": "tur",
+            # Ukrainian
+            "uk": "ukr",
+            # Urdu
+            "ur": "urd",
+            # Northern Uzbek
+            "uz": "uzn",
+            # Vietnamese
+            "vi": "vie",
+            # Yoruba
+            "yo": "yor",
+            # Cantonese
+            "yue": "yue",
+            # Standard Malay
+            "ms": "zsm",
+            # Zulu
+            "zu": "zul",
+        }
+        return seamless_lang_code.get(lang_code, "eng")
+
+    @method()
+    def translate_text(self, text: str, source_language: str, target_language: str):
+        with self.lock:
+            translation_result, _ = self.translator.predict(
+                text,
+                "t2tt",
+                src_lang=self.get_seamless_lang_code(source_language),
+                tgt_lang=self.get_seamless_lang_code(target_language),
+                unit_generation_ngram_filtering=True,
+            )
+        translated_text = str(translation_result[0])
+        return {"text": {source_language: text, target_language: translated_text}}
+
+
+# -------------------------------------------------------------------
+# Web API
+# -------------------------------------------------------------------
+
+
+@app.function(
+    scaledown_window=60,
+    timeout=60,
+    allow_concurrent_inputs=40,
+    secrets=[
+        Secret.from_name("reflector-gpu"),
+    ],
+)
+@asgi_app()
+def web():
+    from fastapi import Body, Depends, FastAPI, HTTPException, status
+    from fastapi.security import OAuth2PasswordBearer
+    from typing_extensions import Annotated
+
+    translatorstub = Translator()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    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 TranslateResponse(BaseModel):
+        result: dict
+
+    @app.post("/translate", dependencies=[Depends(apikey_auth)])
+    async def translate(
+        text: str,
+        source_language: Annotated[str, Body(...)] = "en",
+        target_language: Annotated[str, Body(...)] = "fr",
+    ) -> TranslateResponse:
+        func = translatorstub.translate_text.spawn(
+            text=text,
+            source_language=source_language,
+            target_language=target_language,
+        )
+        result = func.get()
+        return result
+
+    return app
--- a/gpu/self_hosted/.env.example
+++ b/gpu/self_hosted/.env.example
@@ -0,0 +1,2 @@
+REFLECTOR_GPU_APIKEY=
+HF_TOKEN=
--- a/gpu/self_hosted/.gitignore
+++ b/gpu/self_hosted/.gitignore
@@ -0,0 +1,38 @@
+cache/
+
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Env and secrets
+.env
+*.env
+*.secret
+HF_TOKEN
+REFLECTOR_GPU_APIKEY
+
+# Virtual env / uv
+.venv/
+venv/
+ENV/
+uv/
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Logs
+*.log
--- a/gpu/self_hosted/Dockerfile
+++ b/gpu/self_hosted/Dockerfile
@@ -0,0 +1,46 @@
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 \
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
+
+WORKDIR /tmp
+RUN apt-get update \
+ && apt-get install -y \
+    ffmpeg \
+    curl \
+    ca-certificates \
+    gnupg \
+    wget \
+ && apt-get clean
+# Add NVIDIA CUDA repo for Debian 12 (bookworm) and install cuDNN 9 for CUDA 12
+ADD https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb /cuda-keyring.deb
+RUN dpkg -i /cuda-keyring.deb \
+ && rm /cuda-keyring.deb \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+    cuda-cudart-12-6 \
+    libcublas-12-6 \
+    libcudnn9-cuda-12 \
+    libcudnn9-dev-cuda-12 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+ADD https://astral.sh/uv/install.sh /uv-installer.sh
+RUN sh /uv-installer.sh && rm /uv-installer.sh
+ENV PATH="/root/.local/bin/:$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"]
+
+
--- a/gpu/self_hosted/README.md
+++ b/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
--- a/gpu/self_hosted/app/auth.py
+++ b/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"},
+    )
--- a/gpu/self_hosted/app/config.py
+++ b/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")
--- a/gpu/self_hosted/app/factory.py
+++ b/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
--- a/gpu/self_hosted/app/routers/diarization.py
+++ b/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)
--- a/gpu/self_hosted/app/routers/transcription.py
+++ b/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
--- a/gpu/self_hosted/app/routers/translation.py
+++ b/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)
--- a/gpu/self_hosted/app/services/diarizer.py
+++ b/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}
--- a/gpu/self_hosted/app/services/transcriber.py
+++ b/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}
--- a/gpu/self_hosted/app/services/translator.py
+++ b/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}}
--- a/gpu/self_hosted/app/utils.py
+++ b/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)
--- a/gpu/self_hosted/compose.yml
+++ b/gpu/self_hosted/compose.yml
@@ -0,0 +1,10 @@
+services:
+  reflector_gpu:
+    build:
+      context: .
+    ports:
+      - "8000:8000"
+    env_file:
+      - .env
+    volumes:
+      - ./cache:/root/.cache
--- a/gpu/self_hosted/main.py
+++ b/gpu/self_hosted/main.py
@@ -0,0 +1,3 @@
+from app.factory import create_app
+
+app = create_app()
--- a/gpu/self_hosted/pyproject.toml
+++ b/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",
+]
--- a/gpu/self_hosted/runserver.sh
+++ b/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
+
+
--- a/gpu/self_hosted/uv.lock
+++ b/gpu/self_hosted/uv.lock