Update app.py

app.py

@@ -112,8 +112,8 @@ def transcribe_audio(audio_file, chunk_length_s=30):
         attention_mask = torch.ones(inputs.input_features.shape[:2], dtype=torch.long, device=device)
 
         # Check the dimensions and values of the attention mask
-        assert attention_mask.shape == (1, input_features.shape[1]), "Attention mask dimensions do not match the input features."
-        assert (attention_mask.sum().item() == input_features.shape[1]), "Attention mask has incorrect values."
+        assert attention_mask.shape == (2, input_features.shape[2]), "Attention mask dimensions do not match the input features."
+        assert (attention_mask.sum().item() == input_features.shape[2]), "Attention mask has incorrect values."
 
         # ASR model inference on the chunk
         with torch.no_grad():
@@ -124,8 +124,8 @@ def transcribe_audio(audio_file, chunk_length_s=30):
             )
 
             # new processor object with desired configuration
-            new_processor = processor.add_special_tokens({'pad_token': '[PAD]'})
-            chunk_text = new_processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
+            #new_processor = processor.add_special_tokens({'eos_token': '[EOS]'})
+            chunk_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
             full_text.append(chunk_text)
 
     # Combine the transcribed text from all chunks