app.py

import gradio as gr
import torch
from transformers import pipeline
import librosa
import soundfile as sf
import spaces
import os

def split_audio(audio_data, sr, chunk_duration=30):
    """Split audio into chunks of chunk_duration seconds."""
    chunks = []
    for start in range(0, len(audio_data), int(chunk_duration * sr)):
        end = start + int(chunk_duration * sr)
        chunks.append(audio_data[start:end])
    return chunks

def transcribe_long_audio(audio_input, transcriber, chunk_duration=30):
    """Transcribe long audio by splitting into smaller chunks."""
    # Check if audio_input is a file path or raw data
    if isinstance(audio_input, str):
        audio_data, sr = librosa.load(audio_input, sr=None)
    else:  # Raw audio data (e.g., from recording)
        audio_data, sr = audio_input

    chunks = split_audio(audio_data, sr, chunk_duration)
    transcriptions = []
    for i, chunk in enumerate(chunks):
        temp_path = f"temp_chunk_{i}.wav"
        sf.write(temp_path, chunk, sr)  # Save the chunk as a WAV file
        transcription = transcriber(temp_path)["text"]
        transcriptions.append(transcription)
        os.remove(temp_path)  # Clean up temporary files
    return " ".join(transcriptions)

@spaces.GPU(duration=3)
def main():
    # Force GPU if available, fallback to CPU
    device = 0 if torch.cuda.is_available() else -1

    try:
        # Load models with explicit device
        transcriber = pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device)
        summarizer = pipeline("summarization", model="facebook/bart-large-cnn")
    except Exception as e:
        print(f"Error loading models: {e}")
        raise

    # Function to process audio
    def process_audio(audio_input):
        try:
            # Transcribe the audio (long-form support)
            transcription = transcribe_long_audio(audio_input, transcriber, chunk_duration=30)
            # Summarize the transcription
            summary = summarizer(transcription, max_length=50, min_length=10, do_sample=False)[0]["summary_text"]
            return transcription, summary
        except Exception as e:
            return f"Error processing audio: {e}", ""

    # Gradio Interface with Horizontal Layout
    with gr.Blocks() as interface:
        with gr.Row():
            with gr.Column():
                audio_input = gr.Audio(source="microphone", type="numpy", label="Record or Upload Audio")
                process_button = gr.Button("Process Audio")
            with gr.Column():
                transcription_output = gr.Textbox(label="Full Transcription", lines=10)
                summary_output = gr.Textbox(label="Summary", lines=5)

        process_button.click(
            process_audio,
            inputs=[audio_input],
            outputs=[transcription_output, summary_output]
        )

    # Launch the interface with optional public sharing
    interface.launch(share=True)

# Run the main function
if __name__ == "__main__":
    main()