Fix Whisper KV cache KeyError by disabling fp16 and adding fallback

speech_recognizer.py

@@ -27,13 +27,27 @@ class SpeechRecognizer:
         task: str = "transcribe"
     ) -> Dict[str, any]:
         # Transcribe with Whisper (with word-level timestamps for better pause detection)
-        result = self.model.transcribe(
-            audio_path,
-            language=language,
-            task=task,
-            verbose=False,
-            word_timestamps=True
-        )
+        # Using fp16=False to avoid KV cache issues in production environments
+        try:
+            result = self.model.transcribe(
+                audio_path,
+                language=language,
+                task=task,
+                verbose=False,
+                word_timestamps=True,
+                fp16=False  # Disable fp16 to avoid KV cache KeyError
+            )
+        except (KeyError, RuntimeError) as e:
+            # Fallback: transcribe without word timestamps if KV cache fails
+            print(f"Warning: Word-level timestamps failed ({e}), retrying without them...")
+            result = self.model.transcribe(
+                audio_path,
+                language=language,
+                task=task,
+                verbose=False,
+                word_timestamps=False,
+                fp16=False
+            )
         
         transcription = result['text'].strip()
         detected_language = result.get('language', 'unknown')