app.py

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

@spaces.GPU(duration=60)
def transcribe_audio(audio_file):
    try:
        pipe = transformers.pipeline(
            model='sarvamai/shuka_v1',
            trust_remote_code=True,
            device=0,
            torch_dtype=torch.bfloat16
        )

        # Load the audio file at 16kHz
        audio, sr = librosa.load(audio_file, sr=16000)
        
        # Ensure audio is a floating-point numpy array
        audio = np.array(audio, dtype=np.float32)
        # Convert multi-channel audio to mono if needed
        if audio.ndim > 1:
            audio = np.mean(audio, axis=-1)

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

        # Change the conversation turns to instruct transcription
        turns = [
            {'role': 'system', 'content': 'Please transcribe the following audio exactly.'},
            {'role': 'user', 'content': '<|audio|>'}
        ]
        
        # Debug: Print the initial turns
        print(f"Initial turns: {turns}")

        # Call the model with the audio and transcription 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:
        return f"Error: {str(e)}"

iface = gr.Interface(
    fn=transcribe_audio,
    inputs=gr.Audio(sources="microphone", type="filepath"),
    outputs="text",
    title="Shuka ASR Demo",
    description="Speak into your microphone, and the model will transcribe your speech.",
    live=True
)

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