Thonburian Whisper demo with Gradio > 4.0.0

app.py

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

MODEL_NAME = "biodatlab/whisper-th-medium-combined"  # this always needs to stay in line 8 :D sorry for the hackiness
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 is not None) and (file_upload is not None):
        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"
        )
    elif (microphone is None) and (file_upload is None):
        return "ERROR: You have to either use the microphone or upload an audio file"

    file = microphone if microphone is not None else file_upload

    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):
    yt = pt.YouTube(yt_url)
    html_embed_str = _return_yt_html_embed(yt_url)
    stream = yt.streams.filter(only_audio=True)[0]
    stream.download(filename="audio.mp3")

    text = pipe("audio.mp3", generate_kwargs={"language":"<|th|>", "task":"transcribe"}, batch_size=16)["text"]

    return html_embed_str, text

with gr.Blocks(theme=gr.themes.HuggingFace()) as demo:
    gr.Markdown(f"# Whisper Demo Thai 🇹🇭")
    
    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(source="microphone", type="filepath", label="Microphone Input")
                audio_file = gr.Audio(source="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])

demo.launch(enable_queue=True)