Updated Gradio App

app.py

@@ -15,26 +15,33 @@ def transcribe_and_respond(audio_file):
             torch_dtype=torch.bfloat16
         )
 
-        # Load the audio file
+        # Load the audio file at 16kHz
         audio, sr = librosa.load(audio_file, sr=16000)
-
-        # Print audio properties for debugging
+        
+        # Ensure audio is a contiguous floating-point numpy array
+        audio = np.ascontiguousarray(audio, dtype=np.float32)
+        
+        # If audio has more than one channel, convert to mono by averaging the channels
+        if audio.ndim > 1:
+            audio = np.mean(audio, axis=-1)
+        
+        # Debug: Print audio properties
         print(f"Audio dtype: {audio.dtype}, Audio shape: {audio.shape}, Sample rate: {sr}")
-
+        
         turns = [
-            {'role': 'system', 'content': 'Print text.'},
+            {'role': 'system', 'content': 'Please transcribe the following audio exactly.'},
             {'role': 'user', 'content': '<|audio|>'}
         ]
-
+        
         # Debug: Print the initial turns
         print(f"Initial turns: {turns}")
-
+        
         # Call the model with the audio and prompt
         output = pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=512)
-
+        
         # Debug: Print the final output from the model
         print(f"Model output: {output}")
-
+        
         return output
 
     except Exception as e:
@@ -44,8 +51,8 @@ iface = gr.Interface(
     fn=transcribe_and_respond,
     inputs=gr.Audio(sources="microphone", type="filepath"),
     outputs="text",
-    title="Live Transcription",
-    description="Speak into your microphone, and the model will Transcribe.",
+    title="Live Transcription and Response",
+    description="Speak into your microphone, and the model will transcribe your speech.",
     live=True
 )