update

handler.py

@@ -179,7 +179,7 @@ class EndpointHandler():
         
         self.diarize_model = whisperx.DiarizationPipeline( 
                 "pyannote/speaker-diarization-3.0",
-                use_auth_token="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", device="cpu")
+                use_auth_token="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", device=device)
         
         logger.info(f"Model for diarization initialized")
 
@@ -235,6 +235,9 @@ class EndpointHandler():
             if info:
                 print(transcription["segments"][0:10000]) # before alignment
             logger.info(transcription["segments"][0:10000])
+            if len(transcription["segments"]) == 0:
+                logger.warning("No transcription")
+                return {"transcription": transcription["segments"]}
             
             # 3. align
             if alignment:
@@ -262,10 +265,7 @@ class EndpointHandler():
             logger.info(transcription["segments"][0:10000])  # segments are now assigned speaker IDs
 
             if torch.cuda.is_available():
-                logger.info("--------------- GPU ------------------------")
-                logger.info(display_gpu_infos())
-                torch.cuda.empty_cache()
-                logger.info("--------------- GPU AFTER empty_cache ------------------------")
+                logger.info("--------------- GPU AFTER ------------------------")
                 logger.info(display_gpu_infos())
             
             # results_json = json.dumps(results)

requirements.txt

@@ -2,4 +2,5 @@ git+https://github.com/m-bain/whisperX.git
 ffmpeg-python==0.2.0
 setuptools
 torch==2.0.0
-torchaudio==2.0.0
+torchaudio==2.0.0
+ctranslate2==3.22.0