run_demo_hf.py

import logging
import warnings

import gradio as gr
import pytube as pt
import torch
from huggingface_hub import model_info
from transformers import pipeline
from transformers.utils.logging import disable_progress_bar

warnings.filterwarnings("ignore")
disable_progress_bar()

# MODEL_NAME = "bofenghuang/whisper-large-v2-cv11-french"
MODEL_NAME = "bofenghuang/whisper-large-v3-french"
# MODEL_NAME = "/home/bhuang/transformers/examples/pytorch/speech-recognition/outputs/hf_whisper/whisper-large-v3-ft-french-pnc-ep5-bs280-lr4e6-wd001-audioaug-specaug"
# MODEL_NAME = "/home/bhuang/transformers/examples/pytorch/speech-recognition/outputs/hf_whisper/tmp_model"
# MODEL_NAME = "/projects/bhuang/models/asr/public/whisper-large-v3-french"
CHUNK_LENGTH_S = 30

logging.basicConfig(
    format="%(asctime)s [%(levelname)s] [%(name)s] %(message)s",
    datefmt="%Y-%m-%dT%H:%M:%SZ",
)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)

device = 0 if torch.cuda.is_available() else "cpu"
logger.info(f"Model will be loaded on device `{device}`")

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

pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language="fr", 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)["text"]

    logger.info(f"Transcription: {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")["text"]

    logger.info(f'Transcription of "{yt_url}": {text}')

    return html_embed_str, text


demo = gr.Blocks()

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.components.Audio(sources="microphone", type="filepath", label="Record"),
        gr.components.Audio(sources="upload", type="filepath", label="Upload File"),
    ],
    # outputs="text",
    outputs=gr.components.Textbox(label="Transcription", show_copy_button=True),
    # layout="horizontal",
    theme="huggingface",
    title="Whisper French Demo 🇫🇷 : Transcribe Audio",
    # description=(
    #     "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the the fine-tuned"
    #     f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"
    #     " of arbitrary length."
    # ),
    allow_flagging="never",
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[gr.components.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL")],
    # outputs=["html", "text"],
    outputs=[
        gr.components.HTML(label="YouTube Page"),
        gr.components.Textbox(label="Transcription", show_copy_button=True),
    ],
    # layout="horizontal",
    theme="huggingface",
    title="Whisper French Demo 🇫🇷 : Transcribe YouTube",
    # description=(
    #     "Transcribe long-form YouTube videos with the click of a button! Demo uses the the fine-tuned checkpoint:"
    #     f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files of"
    #     " arbitrary length."
    # ),
    allow_flagging="never",
)

with demo:
    gr.TabbedInterface([mf_transcribe, yt_transcribe], ["Transcribe Audio", "Transcribe YouTube"])

# demo.launch(server_name="0.0.0.0", debug=True, share=True)
# demo.launch(enable_queue=True)
# see https://github.com/gradio-app/gradio/issues/2551
demo.queue(max_size=10).launch(server_name="0.0.0.0", debug=True, share=True, ssl_certfile="/home/bhuang/tools/cert.pem", ssl_keyfile="/home/bhuang/tools/key.pem", ssl_verify=False)