app.py

import gradio as gr
import os
import torch
import gc
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM
import tempfile
import yt_dlp

summarizer = pipeline("summarization", model="facebook/bart-large-cnn")
asr_pipeline = pipeline("automatic-speech-recognition", model="openai/whisper-base")

def download_audio(url: str, temp_dir: str) -> str:
    output_path = os.path.join(temp_dir, "audio.%(ext)s")
    ydl_opts = {
        'format': 'bestaudio/best',
        'outtmpl': output_path,
        'quiet': True,
        'postprocessors': [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'mp3',
            'preferredquality': '192',
        }],
    }
    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        ydl.download([url])
    return output_path.replace('%(ext)s', 'mp3')

def process_video(url: str) -> str:
    with tempfile.TemporaryDirectory() as tmpdir:
        audio_path = download_audio(url, tmpdir)
        transcription_result = asr_pipeline(audio_path)
        text = transcription_result['text']
        if len(text.strip()) < 50:
            return "Transcription too short or unclear"
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        summary_result = summarizer(text, max_length=150, min_length=50, do_sample=False)
        return summary_result[0]['summary_text']

def main(url):
    return process_video(url)

iface = gr.Interface(fn=main, inputs="text", outputs="text", title="YouTube Audio Summarizer")
iface.launch()