app.py

import transformers
import gradio as gr
import librosa
import torch
import spaces
import numpy as np

# Maintain conversation history globally
conversation_history = []

@spaces.GPU(duration=15)
def transcribe_and_respond(audio_file, chat_history):
    try:
        pipe = transformers.pipeline(
            model='sarvamai/shuka_v1',
            trust_remote_code=True,
            device=0,
            torch_dtype=torch.bfloat16
        )

        # Load the audio file
        audio, sr = librosa.load(audio_file, sr=16000)

        # Print audio properties for debugging
        print(f"Audio dtype: {audio.dtype}, Audio shape: {audio.shape}, Sample rate: {sr}")

        # Prepare conversation turns
        turns = chat_history.copy()  # Take existing chat history

        # Add the audio to the current turn
        turns.append({'role': 'user', 'content': '<|audio|>'})

        # Debug: Print the updated turns
        print(f"Updated turns: {turns}")

        # Call the model with the updated conversation turns and audio
        output = pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=512)

        # Append the model's response to the chat history
        turns.append({'role': 'system', 'content': output})

        # Debug: Print the model's response
        print(f"Model output: {output}")

        return turns, turns  # Return updated history to display

    except Exception as e:
        return f"Error: {str(e)}"

# Chat interface setup
iface = gr.Interface(
    fn=transcribe_and_respond,
    inputs=[gr.Audio(sources="microphone", type="filepath", label="Your Audio (Microphone)"), gr.Chatbot(label="Conversation")],
    outputs=gr.Chatbot(label="Conversation"),
    title="🎙️ AI Chat with Live Transcription",
    description="Talk to the AI through your microphone, and it will respond conversationally based on the ongoing chat. Keep the conversation going!",
    live=True,
    allow_flagging="auto",
    enable_queue=True
)

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