Updated Gradio App

app.py

@@ -15,33 +15,26 @@ def transcribe_and_respond(audio_file):
             torch_dtype=torch.bfloat16
         )
 
-        # Load the audio file at 16kHz
+        # Load the audio file
         audio, sr = librosa.load(audio_file, sr=16000)
-        
-        # 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 audio properties for debugging
         print(f"Audio dtype: {audio.dtype}, Audio shape: {audio.shape}, Sample rate: {sr}")
-        
+
         turns = [
-            {'role': 'system', 'content': 'Please transcribe the following audio exactly.'},
+            {'role': 'system', 'content': 'Repeat the following text exactly, without any changes'},
             {'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:
@@ -52,9 +45,9 @@ iface = gr.Interface(
     inputs=gr.Audio(sources="microphone", type="filepath"),
     outputs="text",
     title="Live Transcription and Response",
-    description="Speak into your microphone, and the model will transcribe your speech.",
+    description="Speak into your microphone, and the model will respond naturally and informatively.",
     live=True
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()

requirements.txt

@@ -0,0 +1,7 @@
+transformers==4.41.2
+peft==0.11.1
+librosa==0.10.2
+gradio==3.50.2
+huggingface_hub
+torch
+spaces