helvekami/ShukaNote

1import transformers2import gradio as gr3import librosa4import torch5import spaces6import numpy as np7import tempfile8 9@spaces.GPU(duration=60)10def transcribe_and_respond(audio_file, email):11    try:12        pipe = transformers.pipeline(13            model='sarvamai/shuka_v1',14            trust_remote_code=True,15            device=0,16            torch_dtype=torch.bfloat1617        )18 19        # Load the audio file at 16kHz20        audio, sr = librosa.load(audio_file, sr=16000)21        # Convert the audio to a contiguous float32 array22        audio = np.ascontiguousarray(audio, dtype=np.float32)23        # If audio is multi-channel, convert to mono by averaging channels24        if audio.ndim > 1:25            audio = np.mean(audio, axis=-1)26        27        # Debug: Print audio properties28        print(f"Audio dtype: {audio.dtype}, Audio shape: {audio.shape}, Sample rate: {sr}")29        30        # Set up the prompt to get key takeaways31        turns = [32            {'role': 'system', 'content': 'You are an exact echo assistant. Output the previous text exactly as given, without any modifications.'},33            {'role': 'user', 'content': '<|audio|>'}34        ]35        print(f"Initial turns: {turns}")36        37        # Run the model inference (this call is synchronous)38        output = pipe({'audio': audio, 'turns': turns, 'sampling_rate': sr}, max_new_tokens=10000)39        print(f"Model output: {output}")40        41        # Extract transcript text from the output42        transcript = str(output)43        if email and email.strip():44            transcript = f"Email provided: {email}\n\n{transcript}"45        46        # Write the transcript to a temporary file for download47        with tempfile.NamedTemporaryFile(delete=False, mode='w', suffix='.txt') as tmp:48            tmp.write(transcript)49            transcript_file = tmp.name50        51        # Return transcript text and file download path52        return transcript, transcript_file53 54    except Exception as e:55        return f"Error: {str(e)}", ""56 57iface = gr.Interface(58    fn=transcribe_and_respond,59    inputs=[60        gr.Audio(sources=["upload", "microphone"], type="filepath"),61        # gr.Textbox(label="Email", placeholder="Enter your email address (optional)")62    ],63    outputs=[64        gr.Textbox(label="Transcript"),65        gr.File(label="Download Transcript")66    ],67    title="ShukaNotesApp",68    description="Upload or record your meeting audio, and download the transcript."69)70 71if __name__ == "__main__":72    iface.launch()73