1. Removed global pipeline initialization to prevent warm-up crashes

2. Added lazy loading of ASR pipeline
3. Fixed input processing with proper features and attention mask
4. Improved error handling
5. Proper audio preprocessing with correct sampling rate

src/app.py

@@ -41,33 +41,33 @@ MODEL_OPTIONS = {
     }
 }
 
-# Initialize Whisper components
+# Initialize Whisper components globally (these are lightweight)
 feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base.en")
 tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-base.en")
-
-# Configure transcription pipeline with only necessary components
-transcriber = pipeline(
-    "automatic-speech-recognition",
-    model="openai/whisper-base.en",
-    chunk_length_s=30,
-    stride_length_s=5,
-    device="cpu",
-    torch_dtype=torch.float32,
-    generate_kwargs={
-        "use_cache": True,
-        "return_timestamps": True
-    }
-)
+processor = WhisperProcessor(feature_extractor, tokenizer)
+
+def get_asr_pipeline():
+    """Lazy load ASR pipeline with proper configuration."""
+    global transcriber
+    if "transcriber" not in globals():
+        transcriber = pipeline(
+            "automatic-speech-recognition",
+            model="openai/whisper-base.en",
+            chunk_length_s=30,
+            stride_length_s=5,
+            device="cpu",
+            torch_dtype=torch.float32
+        )
+    return transcriber
 
 # Audio preprocessing function
-def prepare_audio_features(audio_array, sample_rate):
-    """Prepare audio features with proper format."""
-    # Convert stereo to mono
+def process_audio(audio_array, sample_rate):
+    """Pre-process audio for Whisper."""
     if audio_array.ndim > 1:
         audio_array = audio_array.mean(axis=1)
-    audio_array = audio_array.astype(np.float32)
     
     # Normalize audio
+    audio_array = audio_array.astype(np.float32)
     audio_array /= np.max(np.abs(audio_array))
     
     # Resample to 16kHz if needed
@@ -77,10 +77,16 @@ def prepare_audio_features(audio_array, sample_rate):
         audio_tensor = resampler(audio_tensor)
         audio_array = audio_tensor.numpy()
     
-    # Return proper dictionary format for pipeline
+    # Process with correct input format
+    inputs = processor(
+        audio_array, 
+        sampling_rate=16000, 
+        return_tensors="pt"
+    )
+    
     return {
-        "raw": audio_array,
-        "sampling_rate": 16000
+        "input_features": inputs.input_features,
+        "attention_mask": inputs.attention_mask
     }
 
 # Update transcriber configuration
@@ -526,14 +532,12 @@ with gr.Blocks(
         
         try:
             sample_rate, audio_array = audio
+            features = process_audio(audio_array, sample_rate)
             
-            # Process audio and get proper format
-            input_features = prepare_audio_features(audio_array, sample_rate)
-            
-            # Pass to transcriber
-            result = transcriber(input_features)
+            # Get pipeline and transcribe
+            asr = get_asr_pipeline()
+            result = asr(features)
             
-            # Extract text from result
             if isinstance(result, dict):
                 return result.get("text", "").strip()
             elif isinstance(result, str):