Docker configuration

.dockerignore

@@ -2,6 +2,8 @@
 /captures/
 /clients/
 /docs/
+local.yaml
+db.sqlite3
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

docker-compose.yml

@@ -4,9 +4,10 @@ services:
   backend:
     image: etown/untitledai:latest
     environment:
-      - UNTITLEDAI_TRANSCRIPTION_HF_TOKEN=${UNTITLEDAI_TRANSCRIPTION_HF_TOKEN}
+      - UNTITLEDAI_ASYNC_WHISPER_HF_TOKEN=${UNTITLEDAI_ASYNC_WHISPER_HF_TOKEN}
       - UNTITLEDAI_USER_CLIENT_TOKEN=${UNTITLEDAI_USER_CLIENT_TOKEN:-change_me}
       - UNTITLEDAI_LLM_API_BASE_URL=${UNTITLEDAI_LLM_API_BASE_URL:-http://host.docker.internal:11434}
+      - OMP_NUM_THREADS=1
 
     shm_size: 3g
     ports:
@@ -21,8 +22,5 @@ services:
       - UNTITLEDAI_USER_CLIENT_TOKEN=${UNTITLEDAI_USER_CLIENT_TOKEN:-change_me}
     ports:
       - "3000:3000"
-    volumes:
-      - ./clients/web:/app
     depends_on:
       - backend
-

untitledai/services/stt/asynchronous/async_whisper/async_whisper_transcription_server.py

@@ -43,15 +43,16 @@ class AsyncWhisperTranscriptionServer:
     def __init__(self, config: AsyncWhisperConfiguration):
         self._config = config
         self.app = FastAPI()
-        self._transcription_model, self._diarize_model, self._verification_model = self._load_models()
+        self._transcription_model, self._diarize_model, self._verification_model, self._alignment_model, self._alignment_metadata = self._load_models()
         self._setup_routes()
 
     def _load_models(self):
         logger.info(f"Transcription model: {self._config.model}")
         transcription_model = whisperx.load_model(self._config.model, self._config.device, compute_type=self._config.compute_type)
         diarize_model = whisperx.DiarizationPipeline(use_auth_token=self._config.hf_token, device=self._config.device)
         verification_model = SpeakerRecognition.from_hparams(source=self._config.verification_model_source, savedir=self._config.verification_model_savedir, run_opts={"device": self._config.device})
-        return transcription_model, diarize_model, verification_model
+        alignment_model, alignment_metadata = whisperx.load_align_model(language_code="en", device=self._config.device)
+        return transcription_model, diarize_model, verification_model, alignment_model, alignment_metadata
 
     def _setup_routes(self):
         # Main endpoint takes a file path and returns a transcription response
@@ -84,8 +85,7 @@ async def _transcribe_audio(self, main_audio_filepath, voice_sample_filepath=Non
         logger.info(f"Initial transcription complete. Total segments: {len(initial_transcription)}")
 
         # Align whisper output
-        model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=self._config.device)
-        result = whisperx.align(initial_transcription, model_a, metadata, audio, device=self._config.device, return_char_alignments=False)
+        result = whisperx.align(initial_transcription, self._alignment_model, self._alignment_metadata, audio, device=self._config.device, return_char_alignments=False)
 
         # Speaker diarization
         diarize_segments = self._diarize_model(audio)

untitledai/services/stt/asynchronous/async_whisper_transcription_service.py

@@ -21,11 +21,11 @@ async def transcribe_audio(self, main_audio_filepath, voice_sample_filepath=None
         url = f"http://{self._config.host}:{self._config.port}/transcribe/"
 
         try:
+            logger.info(f"Sending request to local async whisper server at {url}...")
             response = await self.http_client.post(url, json=payload)
             response.raise_for_status()
             response_string = response.text 
 
-            logger.info(f"Sending request to local async whisper server at {url}...")
             transcript_response = TranscriptionResponse.model_validate_json(response_string)
             utterances = []
             for whisper_utterance in transcript_response.utterances: