app.py

import os
# from pytube import YouTube
from pytubefix import YouTube
from pytubefix.cli import on_progress
import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import gradio as gr


def seconds_to_hms(seconds):
    if type(seconds).__name__ != 'float':
        return seconds

    hours = seconds // 3600
    minutes = (seconds % 3600) // 60
    seconds = seconds % 60
    return f"{int(hours):02}:{int(minutes):02}:{int(seconds):02}"


def chunk_youtube(yt_url):
    yt = YouTube(yt_url, on_progress_callback=on_progress)
    audio = yt.streams.filter(only_audio=True, file_extension='mp4').first()
    audio.download(filename="audio.mp3")
    yield "Stage 1: Audio download DONE"

    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
    torch.cuda.empty_cache()

    model_id = "openai/whisper-large-v3"

    model = AutoModelForSpeechSeq2Seq.from_pretrained(
        model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
    )
    model.to(device)

    processor = AutoProcessor.from_pretrained(model_id)

    pipe = pipeline(
        "automatic-speech-recognition",
        model=model,
        tokenizer=processor.tokenizer,
        feature_extractor=processor.feature_extractor,
        max_new_tokens=128,
        chunk_length_s=30,
        batch_size=16,
        return_timestamps=True,
        torch_dtype=torch_dtype,
        device=device,
    )

    result = pipe("audio.mp3", return_timestamps=True)
    transcript = result["chunks"]
    yield "Stage 2: LLM processing DONE"

    output = ""

    for i in transcript:
        output += str(seconds_to_hms(i["timestamp"][0])) + " - " + str(seconds_to_hms(i["timestamp"][1])) + ":\n" + i[
            "text"] + "\n"

    yield output


iface = gr.Interface(fn=chunk_youtube, inputs="text", outputs="text", live=True)
iface.launch()