app.py

import gradio as gr
import torch
from transformers import pipeline
import librosa
import spaces

def split_audio(audio_path, chunk_duration=30):
    """Split audio into chunks of chunk_duration seconds."""
    audio, sr = librosa.load(audio_path, sr=None)
    chunks = []
    for start in range(0, len(audio), int(chunk_duration * sr)):
        end = start + int(chunk_duration * sr)
        chunks.append(audio[start:end])
    return chunks, sr

def transcribe_long_audio(audio_path, transcriber, chunk_duration=30):
    """Transcribe long audio by splitting into smaller chunks."""
    chunks, sr = split_audio(audio_path, chunk_duration)
    transcriptions = []
    for chunk in chunks:
        temp_path = "temp_chunk.wav"
        librosa.output.write_wav(temp_path, chunk, sr)
        transcription = transcriber(temp_path)["text"]
        transcriptions.append(transcription)
    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_file):
        try:
            # Transcribe the audio (long-form support)
            transcription = transcribe_long_audio(audio_file, 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(type="filepath", label="Upload Audio File")
                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()