Update app.py

app.py

@@ -32,9 +32,8 @@ pipe = pipeline(
     model=model,
     tokenizer=processor.tokenizer,
     feature_extractor=processor.feature_extractor,
-    max_new_tokens=445,  # 448 max_target_positions - 3 special tokens = 445
     chunk_length_s=30,
-    batch_size=16,
+    batch_size=8,  # Reduced batch size for stability
     return_timestamps="word",  # CrisperWhisper provides accurate word-level timestamps
     torch_dtype=torch_dtype,
     device=device,
@@ -75,16 +74,25 @@ def transcribe_audio_chunk(audio_input, task="transcribe", language=None):
     Transcribe a single audio chunk with CrisperWhisper.
     This model is specifically trained for verbatim transcription.
     """
-    generate_kwargs = {
-        "task": task,
-    }
-    
-    if language:
-        generate_kwargs["language"] = language
-    
-    # CrisperWhisper automatically provides verbatim transcription
-    result = pipe(audio_input, generate_kwargs=generate_kwargs)
-    return result
+    try:
+        generate_kwargs = {
+            "task": task,
+        }
+        
+        if language:
+            generate_kwargs["language"] = language
+        
+        # CrisperWhisper automatically provides verbatim transcription
+        result = pipe(audio_input, generate_kwargs=generate_kwargs)
+        return result
+    except Exception as e:
+        # Fallback: try without generate_kwargs if there's a tensor mismatch
+        print(f"Error with generate_kwargs: {e}")
+        try:
+            result = pipe(audio_input)
+            return result
+        except Exception as e2:
+            raise Exception(f"Transcription failed: {str(e2)}")
 
 def transcribe_audio(audio, task="transcribe", return_timestamps=False, language=None, progress=gr.Progress()):
     """