Update app.py

app.py

@@ -83,7 +83,7 @@ def transcribe_audio(audio_file, chunk_length_s=30):
     if waveform.shape[0] > 1:
         waveform = torch.mean(waveform, dim=0, keepdim=True)
 
-    # Resample the audio to 16000 Hz if it’s not already
+    # Resample audio to 16000 Hz if it’s not already
     if sample_rate != 16000:
         resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
         waveform = resampler(waveform)
@@ -109,9 +109,10 @@ def transcribe_audio(audio_file, chunk_length_s=30):
         inputs = processor(chunk_waveform.squeeze(0).numpy(), sampling_rate=sample_rate, return_tensors="pt")
         input_features = inputs.input_features
 
-        # Explicitly set the attention mask
-        attention_mask = torch.ones(input_features.shape, dtype=torch.long)
-
+        # Create attention mask
+        attention_mask = torch.ones(inputs.input_features.shape[:2], dtype=torch.long, device=device)
+        attention_mask[inputs.input_features == processor.pad_token_id] = 0
+        
         # ASR model inference on the chunk
         with torch.no_grad():
             generated_ids = model.generate(
@@ -258,7 +259,7 @@ def save_to_pdf(text, summary):
 iface = gr.Blocks()
 
 with iface:
-    gr.HTML('<img src="https://huggingface.co/spaces/camparchimedes/transcription_app/blob/main/lol.webp" alt="Banner Image" />')
+    gr.HTML('<img src="https://huggingface.co/spaces/camparchimedes/transcription_app/resolve/main/banner_trans.png" alt="Banner Image">')
     gr.Markdown("*Switch Work's JoJo-versjon som webapp for transkribering av lydfiler til norsk skrift. Språkmodell: NbAiLab/nb-whisper-large, Ekstra: oppsummering, pdf-download*")
 
     with gr.Tabs():