app.py

import gradio as gr
import numpy as np
import librosa
from transformers import pipeline
import tempfile
from functools import lru_cache

# Cache the model to avoid reloading on every interaction
@lru_cache(maxsize=1)
def load_model():
    return pipeline(
        model='fixie-ai/ultravox-v0_5-llama-3_2-1b',
        trust_remote_code=True,
        device_map="auto"  # Automatically uses GPU if available
    )

def process_audio(audio_file, user_message):
    try:
        # Load audio (supports file upload or microphone input)
        if isinstance(audio_file, (str, tempfile._TemporaryFileWrapper)):
            audio_path = audio_file.name if hasattr(audio_file, 'name') else audio_file
            audio, sr = librosa.load(audio_path, sr=16000)
        else:  # Handle direct numpy array from microphone
            sr, audio = audio_file
        
        # Initialize conversation
        turns = [
            {
                "role": "system",
                "content": "You are a friendly and helpful AI assistant. Respond conversationally to the user's audio input."
            },
            {
                "role": "user",
                "content": user_message if user_message else "Describe what you heard in the audio."
            }
        ]
        
        # Get model prediction
        pipe = load_model()
        result = pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=100)
        
        return result[-1]["content"]
    
    except Exception as e:
        return f"Error processing audio: {str(e)}"

# Gradio UI
with gr.Blocks(title="UltraVox Audio Assistant") as demo:
    gr.Markdown("## 🎤 UltraVox Audio Assistant")
    gr.Markdown("Upload an audio file or speak via microphone, then ask questions about it.")
    
    with gr.Row():
        audio_input = gr.Audio(
            sources=["upload", "microphone"],
            type="filepath",
            label="Input Audio"
        )
        text_input = gr.Textbox(
            label="Your Question (Optional)",
            placeholder="Ask me about the audio..."
        )
    
    submit_btn = gr.Button("Process")
    output = gr.Textbox(label="AI Response", interactive=False)
    
    submit_btn.click(
        fn=process_audio,
        inputs=[audio_input, text_input],
        outputs=output
    )
    
    gr.Examples(
        examples=[
            ["examples/weather_report.wav", "What's the weather forecast?"],
            ["examples/meeting_notes.mp3", "Summarize the key points"]
        ],
        inputs=[audio_input, text_input]
    )

if __name__ == "__main__":
    demo.launch()