app.py

import gradio as gr
import transformers
import librosa
import torch

# Load the Shuka model pipeline.
pipe = transformers.pipeline(
    model="sarvamai/shuka_v1",
    trust_remote_code=True,
    device=0 if torch.cuda.is_available() else -1,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else None
)

def process_audio(audio):
    """
    Processes the input audio and returns a text response generated by the Shuka model.
    """
    if audio is None:
        return "No audio provided."

    # Gradio returns a tuple (sample_rate, numpy_array)
    sample_rate, audio_data = audio

    # Resample to 16000 Hz if necessary
    if sample_rate != 16000:
        audio_data = librosa.resample(audio_data, orig_sr=sample_rate, target_sr=16000)
        sample_rate = 16000

    # Define conversation turns with a system prompt and a user prompt that signals audio input
    turns = [
        {'role': 'system', 'content': 'Respond naturally and informatively.'},
        {'role': 'user', 'content': '<|audio|>'}
    ]

    # Run the pipeline with the audio input and conversation context
    result = pipe({'audio': audio_data, 'turns': turns, 'sampling_rate': sample_rate}, max_new_tokens=512)
    
    # Extract the generated text response
    if isinstance(result, list) and len(result) > 0:
        response = result[0].get('generated_text', '')
    else:
        response = str(result)
    return response

# Create the Gradio interface without the 'source' parameter.
iface = gr.Interface(
    fn=process_audio,
    inputs=gr.Audio(type="numpy"),
    outputs="text",
    title="Sarvam AI Shuka Voice Demo",
    description="Upload a voice note and get a response using Sarvam AI's Shuka model."
)

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