app.py

import os
import yt_dlp
import torch
import gradio as gr
import pytube as pt
from transformers import pipeline
from huggingface_hub import model_info


# See available models at https://github.com/biodatlab/thonburian-whisper
MODEL_NAME = "biodatlab/distill-whisper-th-large-v3"  # specify the model name here
lang = "th"

device = 0 if torch.cuda.is_available() else "cpu"

pipe = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=30,
    device=device,
)

pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language=lang, task="transcribe")

def transcribe(microphone, file_upload):
    warn_output = ""
    if microphone and file_upload:
        warn_output = (
            "WARNING: You've uploaded an audio file and used the microphone. "
            "The recorded file from the microphone will be used and the uploaded audio will be discarded.\n"
        )
        file = microphone
    elif microphone:
        file = microphone
    elif file_upload:
        file = file_upload
    else:
        return "ERROR: You have to either use the microphone or upload an audio file"

    text = pipe(file, generate_kwargs={"language":"<|th|>", "task":"transcribe"}, batch_size=16)["text"]
    return warn_output + text


def _return_yt_html_embed(yt_url):
    video_id = yt_url.split("?v=")[-1]
    HTML_str = (
        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
        " </center>"
    )
    return HTML_str

def yt_transcribe(yt_url):
    try:
        ydl_opts = {
            'format': 'bestaudio/best',
            'postprocessors': [{
                'key': 'FFmpegExtractAudio',
                'preferredcodec': 'mp3',
                'preferredquality': '192',
            }],
            'outtmpl': 'audio.%(ext)s',
        }
        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
            info = ydl.extract_info(yt_url, download=True)
            video_id = info['id']
        
        html_embed_str = _return_yt_html_embed(video_id)
        
        text = pipe("audio.mp3", generate_kwargs={"language":"<|th|>", "task":"transcribe"}, batch_size=16)["text"]
        
        # Clean up the downloaded file
        os.remove("audio.mp3")
        
        return html_embed_str, text
    except Exception as e:
        return f"Error: {str(e)}", "An error occurred while processing the YouTube video."


with gr.Blocks() as demo:
    gr.Markdown("# Thonburian Whisper Demo 🇹🇭")
    gr.Image(value="thonburian-whisper-logo.png", show_label=False, container=False, width=400)
    
    with gr.Tab("Transcribe Audio"):
        gr.Markdown(
            f"Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the fine-tuned"
            f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"
            f" of arbitrary length."
        )
        with gr.Row():
            with gr.Column():
                audio_mic = gr.Audio(sources=["microphone"], type="filepath", label="Microphone Input")
                audio_file = gr.Audio(sources=["upload"], type="filepath", label="Audio File Upload")
            with gr.Column():
                text_output = gr.Textbox(label="Transcription Output")
        transcribe_btn = gr.Button("Transcribe")
        transcribe_btn.click(fn=transcribe, inputs=[audio_mic, audio_file], outputs=text_output)
    
    with gr.Tab("Transcribe YouTube"):
        gr.Markdown(
            f"Transcribe long-form YouTube videos with the click of a button! Demo uses the fine-tuned checkpoint:"
            f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files of"
            f" arbitrary length."
        )
        with gr.Row():
            with gr.Column():
                yt_url_input = gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")
            with gr.Column():
                yt_html_output = gr.HTML(label="Video")
                yt_text_output = gr.Textbox(label="Transcription Output")
        yt_transcribe_btn = gr.Button("Transcribe YouTube Video")
        yt_transcribe_btn.click(fn=yt_transcribe, inputs=yt_url_input, outputs=[yt_html_output, yt_text_output])


if __name__ == "__main__":
    demo.queue().launch()