feat: 3-mode selfhosted refactoring (--gpu, --cpu, --hosted) + audio token auth fallback (#896)

* fix: local processing instead of http server for cpu

* add fallback token if service worker doesnt work

* chore: rename processors to keep processor pattern up to date and allow other processors to be createed and used with env vars

server/reflector/auth/__init__.py

@@ -14,6 +14,7 @@ current_user = auth_module.current_user
 current_user_optional = auth_module.current_user_optional
 parse_ws_bearer_token = auth_module.parse_ws_bearer_token
 current_user_ws_optional = auth_module.current_user_ws_optional
+verify_raw_token = auth_module.verify_raw_token
 
 # Optional router (e.g. for /auth/login in password backend)
 router = getattr(auth_module, "router", None)

server/reflector/auth/auth_jwt.py

@@ -144,3 +144,8 @@ async def current_user_ws_optional(websocket: "WebSocket") -> Optional[UserInfo]
     if not token:
         return None
     return await _authenticate_user(token, None, JWTAuth())
+
+
+async def verify_raw_token(token: str) -> Optional[UserInfo]:
+    """Verify a raw JWT token string (used for query-param auth fallback)."""
+    return await _authenticate_user(token, None, JWTAuth())

server/reflector/auth/auth_none.py

@@ -27,3 +27,8 @@ def parse_ws_bearer_token(websocket):
 
 async def current_user_ws_optional(websocket):
     return None
+
+
+async def verify_raw_token(token):
+    """Verify a raw JWT token string (used for query-param auth fallback)."""
+    return None

server/reflector/auth/auth_password.py

@@ -168,6 +168,11 @@ async def current_user_ws_optional(websocket: "WebSocket") -> Optional[UserInfo]
     return await _authenticate_user(token, None)
 
 
+async def verify_raw_token(token: str) -> Optional[UserInfo]:
+    """Verify a raw JWT token string (used for query-param auth fallback)."""
+    return await _authenticate_user(token, None)
+
+
 # --- Login router ---
 router = APIRouter(prefix="/auth", tags=["auth"])
 

server/reflector/processors/__init__.py

@@ -4,6 +4,8 @@ from .audio_diarization_auto import AudioDiarizationAutoProcessor  # noqa: F401
 from .audio_downscale import AudioDownscaleProcessor  # noqa: F401
 from .audio_file_writer import AudioFileWriterProcessor  # noqa: F401
 from .audio_merge import AudioMergeProcessor  # noqa: F401
+from .audio_padding import AudioPaddingProcessor  # noqa: F401
+from .audio_padding_auto import AudioPaddingAutoProcessor  # noqa: F401
 from .audio_transcript import AudioTranscriptProcessor  # noqa: F401
 from .audio_transcript_auto import AudioTranscriptAutoProcessor  # noqa: F401
 from .base import (  # noqa: F401

server/reflector/processors/_audio_download.py

@@ -0,0 +1,86 @@
+"""
+Shared audio download utility for local processors.
+
+Downloads audio from a URL to a temporary file for in-process ML inference.
+"""
+
+import asyncio
+import os
+import tempfile
+from pathlib import Path
+
+import requests
+
+from reflector.logger import logger
+
+S3_TIMEOUT = 60
+
+
+async def download_audio_to_temp(url: str) -> Path:
+    """Download audio from URL to a temporary file.
+
+    The caller is responsible for deleting the temp file after use.
+
+    Args:
+        url: Presigned URL or public URL to download audio from.
+
+    Returns:
+        Path to the downloaded temporary file.
+    """
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(None, _download_blocking, url)
+
+
+def _download_blocking(url: str) -> Path:
+    """Blocking download implementation."""
+    log = logger.bind(url=url[:80])
+    log.info("Downloading audio to temp file")
+
+    response = requests.get(url, stream=True, timeout=S3_TIMEOUT)
+    response.raise_for_status()
+
+    # Determine extension from content-type or URL
+    ext = _detect_extension(url, response.headers.get("content-type", ""))
+
+    fd, tmp_path = tempfile.mkstemp(suffix=ext)
+    try:
+        total_bytes = 0
+        with os.fdopen(fd, "wb") as f:
+            for chunk in response.iter_content(chunk_size=8192):
+                if chunk:
+                    f.write(chunk)
+                    total_bytes += len(chunk)
+        log.info("Audio downloaded", bytes=total_bytes, path=tmp_path)
+        return Path(tmp_path)
+    except Exception:
+        # Clean up on failure
+        try:
+            os.unlink(tmp_path)
+        except OSError:
+            pass
+        raise
+
+
+def _detect_extension(url: str, content_type: str) -> str:
+    """Detect audio file extension from URL or content-type."""
+    # Try URL path first
+    path = url.split("?")[0]  # Strip query params
+    for ext in (".wav", ".mp3", ".mp4", ".m4a", ".webm", ".ogg", ".flac"):
+        if path.lower().endswith(ext):
+            return ext
+
+    # Try content-type
+    ct_map = {
+        "audio/wav": ".wav",
+        "audio/x-wav": ".wav",
+        "audio/mpeg": ".mp3",
+        "audio/mp4": ".m4a",
+        "audio/webm": ".webm",
+        "audio/ogg": ".ogg",
+        "audio/flac": ".flac",
+    }
+    for ct, ext in ct_map.items():
+        if ct in content_type.lower():
+            return ext
+
+    return ".audio"

server/reflector/processors/_marian_translator_service.py

@@ -0,0 +1,76 @@
+"""
+MarianMT translation service.
+
+Singleton service that loads HuggingFace MarianMT translation models
+and reuses them across all MarianMT translator processor instances.
+
+Ported from gpu/self_hosted/app/services/translator.py for in-process use.
+"""
+
+import logging
+import threading
+
+from transformers import MarianMTModel, MarianTokenizer, pipeline
+
+logger = logging.getLogger(__name__)
+
+
+class MarianTranslatorService:
+    """MarianMT text translation service for in-process use."""
+
+    def __init__(self):
+        self._pipeline = None
+        self._current_pair = None
+        self._lock = threading.Lock()
+
+    def load(self, source_language: str = "en", target_language: str = "fr"):
+        """Load the translation model for a specific language pair."""
+        model_name = self._resolve_model_name(source_language, target_language)
+        logger.info(
+            "Loading MarianMT model: %s (%s -> %s)",
+            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)
+        self._current_pair = (source_language.lower(), target_language.lower())
+
+    def _resolve_model_name(self, src: str, tgt: str) -> str:
+        """Resolve language pair to MarianMT model name."""
+        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:
+        """Translate text between languages.
+
+        Args:
+            text: Text to translate.
+            source_language: Source language code (e.g. "en").
+            target_language: Target language code (e.g. "fr").
+
+        Returns:
+            dict with "text" key containing {source_language: original, target_language: translated}.
+        """
+        pair = (source_language.lower(), target_language.lower())
+        if self._pipeline is None or self._current_pair != pair:
+            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}}
+
+
+# Module-level singleton — shared across all MarianMT translator processors
+translator_service = MarianTranslatorService()

server/reflector/processors/_pyannote_diarization_service.py

@@ -0,0 +1,133 @@
+"""
+Pyannote diarization service using pyannote.audio.
+
+Singleton service that loads the pyannote speaker diarization model once
+and reuses it across all pyannote diarization processor instances.
+
+Ported from gpu/self_hosted/app/services/diarizer.py for in-process use.
+"""
+
+import logging
+import tarfile
+import threading
+from pathlib import Path
+from urllib.request import urlopen
+
+import torch
+import torchaudio
+import yaml
+from pyannote.audio import Pipeline
+
+from reflector.settings import settings
+
+logger = logging.getLogger(__name__)
+
+S3_BUNDLE_URL = "https://reflector-public.s3.us-east-1.amazonaws.com/pyannote-speaker-diarization-3.1.tar.gz"
+BUNDLE_CACHE_DIR = Path.home() / ".cache" / "pyannote-bundle"
+
+
+def _ensure_model(cache_dir: Path) -> str:
+    """Download and extract S3 model bundle if not cached."""
+    model_dir = cache_dir / "pyannote-speaker-diarization-3.1"
+    config_path = model_dir / "config.yaml"
+
+    if config_path.exists():
+        logger.info("Using cached model bundle at %s", model_dir)
+        return str(model_dir)
+
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    tarball_path = cache_dir / "model.tar.gz"
+
+    logger.info("Downloading model bundle from %s", S3_BUNDLE_URL)
+    with urlopen(S3_BUNDLE_URL) as response, open(tarball_path, "wb") as f:
+        while chunk := response.read(8192):
+            f.write(chunk)
+
+    logger.info("Extracting model bundle")
+    with tarfile.open(tarball_path, "r:gz") as tar:
+        tar.extractall(path=cache_dir, filter="data")
+    tarball_path.unlink()
+
+    _patch_config(model_dir, cache_dir)
+    return str(model_dir)
+
+
+def _patch_config(model_dir: Path, cache_dir: Path) -> None:
+    """Rewrite config.yaml to reference local pytorch_model.bin paths."""
+    config_path = model_dir / "config.yaml"
+    with open(config_path) as f:
+        config = yaml.safe_load(f)
+
+    config["pipeline"]["params"]["segmentation"] = str(
+        cache_dir / "pyannote-segmentation-3.0" / "pytorch_model.bin"
+    )
+    config["pipeline"]["params"]["embedding"] = str(
+        cache_dir / "pyannote-wespeaker-voxceleb-resnet34-LM" / "pytorch_model.bin"
+    )
+
+    with open(config_path, "w") as f:
+        yaml.dump(config, f)
+
+    logger.info("Patched config.yaml with local model paths")
+
+
+class PyannoteDiarizationService:
+    """Pyannote speaker diarization service for in-process use."""
+
+    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"
+        hf_token = settings.HF_TOKEN
+
+        if hf_token:
+            logger.info("Loading pyannote model from HuggingFace (HF_TOKEN set)")
+            self._pipeline = Pipeline.from_pretrained(
+                "pyannote/speaker-diarization-3.1",
+                use_auth_token=hf_token,
+            )
+        else:
+            logger.info("HF_TOKEN not set — loading model from S3 bundle")
+            model_path = _ensure_model(BUNDLE_CACHE_DIR)
+            config_path = Path(model_path) / "config.yaml"
+            self._pipeline = Pipeline.from_pretrained(str(config_path))
+
+        self._pipeline.to(torch.device(self._device))
+
+    def diarize_file(self, file_path: str, timestamp: float = 0.0) -> dict:
+        """Run speaker diarization on an audio file.
+
+        Args:
+            file_path: Path to the audio file.
+            timestamp: Offset to add to all segment timestamps.
+
+        Returns:
+            dict with "diarization" key containing list of
+            {"start": float, "end": float, "speaker": int} segments.
+        """
+        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}
+            )
+        segments = []
+        for diarization_segment, _, speaker in diarization.itertracks(yield_label=True):
+            segments.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": segments}
+
+
+# Module-level singleton — shared across all pyannote diarization processors
+diarization_service = PyannoteDiarizationService()

server/reflector/processors/audio_diarization_pyannote.py

@@ -0,0 +1,37 @@
+"""
+Pyannote audio diarization processor using pyannote.audio in-process.
+
+Downloads audio from URL, runs pyannote diarization locally,
+and returns speaker segments. No HTTP backend needed.
+"""
+
+import asyncio
+import os
+
+from reflector.processors._audio_download import download_audio_to_temp
+from reflector.processors._pyannote_diarization_service import diarization_service
+from reflector.processors.audio_diarization import AudioDiarizationProcessor
+from reflector.processors.audio_diarization_auto import AudioDiarizationAutoProcessor
+from reflector.processors.types import AudioDiarizationInput
+
+
+class AudioDiarizationPyannoteProcessor(AudioDiarizationProcessor):
+    INPUT_TYPE = AudioDiarizationInput
+
+    async def _diarize(self, data: AudioDiarizationInput):
+        """Run pyannote diarization on audio from URL."""
+        tmp_path = await download_audio_to_temp(data.audio_url)
+        try:
+            loop = asyncio.get_event_loop()
+            result = await loop.run_in_executor(
+                None, diarization_service.diarize_file, str(tmp_path)
+            )
+            return result["diarization"]
+        finally:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+
+
+AudioDiarizationAutoProcessor.register("pyannote", AudioDiarizationPyannoteProcessor)

server/reflector/processors/audio_padding.py

@@ -0,0 +1,23 @@
+"""
+Base class for audio padding processors.
+"""
+
+from pydantic import BaseModel
+
+
+class PaddingResponse(BaseModel):
+    size: int
+    cancelled: bool = False
+
+
+class AudioPaddingProcessor:
+    """Base class for audio padding processors."""
+
+    async def pad_track(
+        self,
+        track_url: str,
+        output_url: str,
+        start_time_seconds: float,
+        track_index: int,
+    ) -> PaddingResponse:
+        raise NotImplementedError

server/reflector/processors/audio_padding_auto.py

@@ -1,9 +1,10 @@
 import importlib
 
+from reflector.processors.audio_padding import AudioPaddingProcessor
 from reflector.settings import settings
 
 
-class AudioPaddingAutoProcessor:
+class AudioPaddingAutoProcessor(AudioPaddingProcessor):
     _registry = {}
 
     @classmethod

server/reflector/processors/audio_padding_modal.py

@@ -6,19 +6,14 @@ import asyncio
 import os
 
 import httpx
-from pydantic import BaseModel
 
 from reflector.hatchet.constants import TIMEOUT_AUDIO
 from reflector.logger import logger
+from reflector.processors.audio_padding import AudioPaddingProcessor, PaddingResponse
 from reflector.processors.audio_padding_auto import AudioPaddingAutoProcessor
 
 
-class PaddingResponse(BaseModel):
-    size: int
-    cancelled: bool = False
-
-
-class AudioPaddingModalProcessor:
+class AudioPaddingModalProcessor(AudioPaddingProcessor):
     """Audio padding processor using Modal.com CPU backend via HTTP."""
 
     def __init__(

server/reflector/processors/audio_padding_pyav.py

@@ -1,5 +1,5 @@
 """
-Local audio padding processor using PyAV.
+PyAV audio padding processor.
 
 Pads audio tracks with silence directly in-process (no HTTP).
 Reuses the shared PyAV utilities from reflector.utils.audio_padding.
@@ -12,15 +12,15 @@ import tempfile
 import av
 
 from reflector.logger import logger
+from reflector.processors.audio_padding import AudioPaddingProcessor, PaddingResponse
 from reflector.processors.audio_padding_auto import AudioPaddingAutoProcessor
-from reflector.processors.audio_padding_modal import PaddingResponse
 from reflector.utils.audio_padding import apply_audio_padding_to_file
 
 S3_TIMEOUT = 60
 
 
-class AudioPaddingLocalProcessor:
-    """Audio padding processor using local PyAV (no HTTP backend)."""
+class AudioPaddingPyavProcessor(AudioPaddingProcessor):
+    """Audio padding processor using PyAV (no HTTP backend)."""
 
     async def pad_track(
         self,
@@ -29,7 +29,7 @@ class AudioPaddingLocalProcessor:
         start_time_seconds: float,
         track_index: int,
     ) -> PaddingResponse:
-        """Pad audio track with silence locally via PyAV.
+        """Pad audio track with silence via PyAV.
 
         Args:
             track_url: Presigned GET URL for source audio track
@@ -130,4 +130,4 @@ class AudioPaddingLocalProcessor:
                 log.warning("Failed to cleanup temp directory", error=str(e))
 
 
-AudioPaddingAutoProcessor.register("local", AudioPaddingLocalProcessor)
+AudioPaddingAutoProcessor.register("pyav", AudioPaddingPyavProcessor)

server/reflector/processors/audio_transcript_whisper.py

@@ -3,13 +3,17 @@ from faster_whisper import WhisperModel
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
 from reflector.processors.audio_transcript_auto import AudioTranscriptAutoProcessor
 from reflector.processors.types import AudioFile, Transcript, Word
+from reflector.settings import settings
 
 
 class AudioTranscriptWhisperProcessor(AudioTranscriptProcessor):
     def __init__(self):
         super().__init__()
         self.model = WhisperModel(
-            "tiny", device="cpu", compute_type="float32", num_workers=12
+            settings.WHISPER_CHUNK_MODEL,
+            device="cpu",
+            compute_type="float32",
+            num_workers=12,
         )
 
     async def _transcript(self, data: AudioFile):

server/reflector/processors/file_diarization_pyannote.py

@@ -0,0 +1,39 @@
+"""
+Pyannote file diarization processor using pyannote.audio in-process.
+
+Downloads audio from URL, runs pyannote diarization locally,
+and returns speaker segments. No HTTP backend needed.
+"""
+
+import asyncio
+import os
+
+from reflector.processors._audio_download import download_audio_to_temp
+from reflector.processors._pyannote_diarization_service import diarization_service
+from reflector.processors.file_diarization import (
+    FileDiarizationInput,
+    FileDiarizationOutput,
+    FileDiarizationProcessor,
+)
+from reflector.processors.file_diarization_auto import FileDiarizationAutoProcessor
+
+
+class FileDiarizationPyannoteProcessor(FileDiarizationProcessor):
+    async def _diarize(self, data: FileDiarizationInput):
+        """Run pyannote diarization on file from URL."""
+        self.logger.info(f"Starting pyannote diarization from {data.audio_url}")
+        tmp_path = await download_audio_to_temp(data.audio_url)
+        try:
+            loop = asyncio.get_event_loop()
+            result = await loop.run_in_executor(
+                None, diarization_service.diarize_file, str(tmp_path)
+            )
+            return FileDiarizationOutput(diarization=result["diarization"])
+        finally:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+
+
+FileDiarizationAutoProcessor.register("pyannote", FileDiarizationPyannoteProcessor)

server/reflector/processors/file_transcript_whisper.py

@@ -0,0 +1,275 @@
+"""
+Local file transcription processor using faster-whisper with Silero VAD pipeline.
+
+Downloads audio from URL, segments it using Silero VAD, transcribes each
+segment with faster-whisper, and merges results. No HTTP backend needed.
+
+VAD pipeline ported from gpu/self_hosted/app/services/transcriber.py.
+"""
+
+import asyncio
+import os
+import shutil
+import subprocess
+import threading
+from typing import Generator
+
+import numpy as np
+from silero_vad import VADIterator, load_silero_vad
+
+from reflector.processors._audio_download import download_audio_to_temp
+from reflector.processors.file_transcript import (
+    FileTranscriptInput,
+    FileTranscriptProcessor,
+)
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.types import Transcript, Word
+from reflector.settings import settings
+
+SAMPLE_RATE = 16000
+
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+
+class FileTranscriptWhisperProcessor(FileTranscriptProcessor):
+    """Transcribe complete audio files using local faster-whisper with VAD."""
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._model = None
+        self._lock = threading.Lock()
+
+    def _ensure_model(self):
+        """Lazy-load the whisper model on first use."""
+        if self._model is not None:
+            return
+
+        import faster_whisper
+        import torch
+
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        compute_type = "float16" if device == "cuda" else "int8"
+        model_name = settings.WHISPER_FILE_MODEL
+
+        self.logger.info(
+            "Loading whisper model",
+            model=model_name,
+            device=device,
+            compute_type=compute_type,
+        )
+        self._model = faster_whisper.WhisperModel(
+            model_name,
+            device=device,
+            compute_type=compute_type,
+            num_workers=1,
+        )
+
+    async def _transcript(self, data: FileTranscriptInput):
+        """Download file, run VAD segmentation, transcribe each segment."""
+        tmp_path = await download_audio_to_temp(data.audio_url)
+        try:
+            loop = asyncio.get_event_loop()
+            result = await loop.run_in_executor(
+                None,
+                self._transcribe_file_blocking,
+                str(tmp_path),
+                data.language,
+            )
+            return result
+        finally:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+
+    def _transcribe_file_blocking(self, file_path: str, language: str) -> Transcript:
+        """Blocking transcription with VAD pipeline."""
+        self._ensure_model()
+
+        audio_array = _load_audio_via_ffmpeg(file_path, SAMPLE_RATE)
+
+        # VAD segmentation → batch merging
+        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))
+
+        # If no speech detected, try transcribing the whole file
+        if not merged_batches:
+            return self._transcribe_whole_file(file_path, language)
+
+        # Transcribe each batch
+        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 = _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)
+
+            for seg in segments:
+                for w in seg.words:
+                    all_words.append(
+                        {
+                            "word": w.word,
+                            "start": round(float(w.start) + start_time, 2),
+                            "end": round(float(w.end) + start_time, 2),
+                        }
+                    )
+
+        all_words = _enforce_word_timing_constraints(all_words)
+
+        words = [
+            Word(text=w["word"], start=w["start"], end=w["end"]) for w in all_words
+        ]
+        words.sort(key=lambda w: w.start)
+        return Transcript(words=words)
+
+    def _transcribe_whole_file(self, file_path: str, language: str) -> Transcript:
+        """Fallback: transcribe entire file without VAD segmentation."""
+        with self._lock:
+            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)
+
+        words = []
+        for seg in segments:
+            for w in seg.words:
+                words.append(
+                    Word(
+                        text=w.word,
+                        start=round(float(w.start), 2),
+                        end=round(float(w.end), 2),
+                    )
+                )
+        return Transcript(words=words)
+
+
+# --- VAD helpers (ported from gpu/self_hosted/app/services/transcriber.py) ---
+# IMPORTANT: This VAD segment logic is duplicated for deployment isolation.
+# If you modify this, consider updating the GPU service copy as well:
+#   - gpu/self_hosted/app/services/transcriber.py
+#   - gpu/modal_deployments/reflector_transcriber.py
+#   - gpu/modal_deployments/reflector_transcriber_parakeet.py
+
+
+def _load_audio_via_ffmpeg(
+    input_path: str, sample_rate: int = SAMPLE_RATE
+) -> np.ndarray:
+    """Load audio file via ffmpeg, converting to mono float32 at target sample rate."""
+    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",
+    ]
+    proc = subprocess.run(
+        cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True
+    )
+    return np.frombuffer(proc.stdout, dtype=np.float32)
+
+
+def _vad_segments(
+    audio_array: np.ndarray,
+    sample_rate: int = SAMPLE_RATE,
+    window_size: int = VAD_CONFIG["window_size"],
+) -> Generator[tuple[float, float], None, None]:
+    """Detect speech segments using Silero VAD."""
+    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
+
+    # Handle case where audio ends while speech is still active
+    if start is not None:
+        audio_duration = len(audio_array) / float(sample_rate)
+        yield (start / float(SAMPLE_RATE), audio_duration)
+
+    iterator.reset_states()
+
+
+def _pad_audio(audio_array: np.ndarray, sample_rate: int = SAMPLE_RATE) -> np.ndarray:
+    """Pad short audio with silence for VAD compatibility."""
+    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(words: list[dict]) -> list[dict]:
+    """Ensure no word end time exceeds the next word's start time."""
+    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
+
+
+FileTranscriptAutoProcessor.register("whisper", FileTranscriptWhisperProcessor)

server/reflector/processors/transcript_translator_marian.py

@@ -0,0 +1,50 @@
+"""
+MarianMT transcript translator processor using HuggingFace MarianMT in-process.
+
+Translates transcript text using HuggingFace MarianMT models
+locally. No HTTP backend needed.
+"""
+
+import asyncio
+
+from reflector.processors._marian_translator_service import translator_service
+from reflector.processors.transcript_translator import TranscriptTranslatorProcessor
+from reflector.processors.transcript_translator_auto import (
+    TranscriptTranslatorAutoProcessor,
+)
+from reflector.processors.types import TranslationLanguages
+
+
+class TranscriptTranslatorMarianProcessor(TranscriptTranslatorProcessor):
+    """Translate transcript text using MarianMT models."""
+
+    async def _translate(self, text: str) -> str | None:
+        source_language = self.get_pref("audio:source_language", "en")
+        target_language = self.get_pref("audio:target_language", "en")
+
+        languages = TranslationLanguages()
+        assert languages.is_supported(target_language)
+
+        self.logger.debug(f"MarianMT translate {text=}")
+
+        loop = asyncio.get_event_loop()
+        result = await loop.run_in_executor(
+            None,
+            translator_service.translate,
+            text,
+            source_language,
+            target_language,
+        )
+
+        if target_language in result["text"]:
+            translation = result["text"][target_language]
+        else:
+            translation = None
+
+        self.logger.debug(f"Translation result: {text=}, {translation=}")
+        return translation
+
+
+TranscriptTranslatorAutoProcessor.register(
+    "marian", TranscriptTranslatorMarianProcessor
+)

server/reflector/settings.py

@@ -40,11 +40,19 @@ class Settings(BaseSettings):
     # backends: silero, frames
     AUDIO_CHUNKER_BACKEND: str = "frames"
 
+    # HuggingFace token for gated models (pyannote diarization in --cpu mode)
+    HF_TOKEN: str | None = None
+
     # Audio Transcription
     # backends:
     #   - whisper: in-process model loading (no HTTP, runs in same process)
     #   - modal: HTTP API client (works with Modal.com OR self-hosted gpu/self_hosted/)
     TRANSCRIPT_BACKEND: str = "whisper"
+
+    # Whisper model sizes for local transcription
+    # Options: "tiny", "base", "small", "medium", "large-v2"
+    WHISPER_CHUNK_MODEL: str = "tiny"
+    WHISPER_FILE_MODEL: str = "tiny"
     TRANSCRIPT_URL: str | None = None
     TRANSCRIPT_TIMEOUT: int = 90
     TRANSCRIPT_FILE_TIMEOUT: int = 600
@@ -100,7 +108,7 @@ class Settings(BaseSettings):
     )
 
     # Diarization
-    # backend: modal — HTTP API client (works with Modal.com OR self-hosted gpu/self_hosted/)
+    # backends: modal — HTTP API client, pyannote — in-process pyannote.audio
     DIARIZATION_ENABLED: bool = True
     DIARIZATION_BACKEND: str = "modal"
     DIARIZATION_URL: str | None = None
@@ -111,9 +119,9 @@ class Settings(BaseSettings):
 
     # Audio Padding
     # backends:
-    #   - local: in-process PyAV padding (no HTTP, runs in same process)
+    #   - pyav: in-process PyAV padding (no HTTP, runs in same process)
     #   - modal: HTTP API client (works with Modal.com OR self-hosted gpu/self_hosted/)
-    PADDING_BACKEND: str = "local"
+    PADDING_BACKEND: str = "pyav"
     PADDING_URL: str | None = None
     PADDING_MODAL_API_KEY: str | None = None
 

server/reflector/views/transcripts_audio.py

@@ -13,7 +13,6 @@ from fastapi import APIRouter, Depends, HTTPException, Request, Response, status
 import reflector.auth as auth
 from reflector.db.transcripts import AudioWaveform, transcripts_controller
 from reflector.settings import settings
-from reflector.views.transcripts import ALGORITHM
 
 from ._range_requests_response import range_requests_response
 
@@ -36,16 +35,23 @@ async def transcript_get_audio_mp3(
 ):
     user_id = user["sub"] if user else None
     if not user_id and token:
-        unauthorized_exception = HTTPException(
-            status_code=status.HTTP_401_UNAUTHORIZED,
-            detail="Invalid or expired token",
-            headers={"WWW-Authenticate": "Bearer"},
-        )
         try:
-            payload = jwt.decode(token, settings.SECRET_KEY, algorithms=[ALGORITHM])
-            user_id: str = payload.get("sub")
-        except jwt.PyJWTError:
-            raise unauthorized_exception
+            token_user = await auth.verify_raw_token(token)
+        except Exception:
+            token_user = None
+        # Fallback: try as internal HS256 token (created by _generate_local_audio_link)
+        if not token_user:
+            try:
+                payload = jwt.decode(token, settings.SECRET_KEY, algorithms=["HS256"])
+                user_id = payload.get("sub")
+            except jwt.PyJWTError:
+                raise HTTPException(
+                    status_code=status.HTTP_401_UNAUTHORIZED,
+                    detail="Invalid or expired token",
+                    headers={"WWW-Authenticate": "Bearer"},
+                )
+        else:
+            user_id = token_user["sub"]
 
     transcript = await transcripts_controller.get_by_id_for_http(
         transcript_id, user_id=user_id