app.py

import transformers
import librosa
import gradio as gr
import spaces

# Load the model pipeline on GPU:0
pipe = transformers.pipeline(
    model='sarvamai/shuka_v1',
    trust_remote_code=True,
    device=0,
    torch_dtype='bfloat16'
)

@spaces.GPU(duration=120)
def transcribe_and_respond(audio_file):
    try:
        # Check if the audio file is valid and exists
        if audio_file is None or not isinstance(audio_file, str):
            raise ValueError("Invalid audio file input.")
        
        # Load the audio using librosa
        audio, sr = librosa.load(audio_file, sr=16000)
        
        # Prepare the conversation turns
        turns = [
            {'role': 'system', 'content': 'Respond naturally and informatively.'},
            {'role': 'user', 'content': ''}
        ]
        
        # Run inference with the pipeline
        response = pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=512)
        
        return response
    except Exception as e:
        return f"Error processing audio: {str(e)}"

# Create the Gradio interface with microphone input
iface = gr.Interface(
    fn=transcribe_and_respond,
    inputs=gr.Audio(sources="microphone", type="filepath"),  # Use the microphone for audio input
    outputs="text",  # The output will be a text response
    title="Voice Input for Transcription and Response",
    description="Record your voice, and the model will respond naturally and informatively."
)

# Launch the Gradio app
iface.launch()