Elizabethx/whisper

1import torch2 3import gradio as gr4import yt_dlp as youtube_dl5from transformers import pipeline6from transformers.pipelines.audio_utils import ffmpeg_read7 8import tempfile9import os10 11MODEL_NAME = "openai/whisper-large-v3"12BATCH_SIZE = 813FILE_LIMIT_MB = 100014YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files15 16device = 0 if torch.cuda.is_available() else "cpu"17 18pipe = pipeline(19    task="automatic-speech-recognition",20    model=MODEL_NAME,21    chunk_length_s=30,22    device=device,23)24 25 26def transcribe(inputs, task):27    if inputs is None:28        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")29 30    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]31    return  text32 33 34def _return_yt_html_embed(yt_url):35    video_id = yt_url.split("?v=")[-1]36    HTML_str = (37        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'38        " </center>"39    )40    return HTML_str41 42def download_yt_audio(yt_url, filename):43    info_loader = youtube_dl.YoutubeDL()44    45    try:46        info = info_loader.extract_info(yt_url, download=False)47    except youtube_dl.utils.DownloadError as err:48        raise gr.Error(str(err))49    50    file_length = info["duration_string"]51    file_h_m_s = file_length.split(":")52    file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]53    54    if len(file_h_m_s) == 1:55        file_h_m_s.insert(0, 0)56    if len(file_h_m_s) == 2:57        file_h_m_s.insert(0, 0)58    file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]59    60    if file_length_s > YT_LENGTH_LIMIT_S:61        yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))62        file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))63        raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")64    65    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}66    67    with youtube_dl.YoutubeDL(ydl_opts) as ydl:68        try:69            ydl.download([yt_url])70        except youtube_dl.utils.ExtractorError as err:71            raise gr.Error(str(err))72 73 74def yt_transcribe(yt_url, task, max_filesize=75.0):75    html_embed_str = _return_yt_html_embed(yt_url)76 77    with tempfile.TemporaryDirectory() as tmpdirname:78        filepath = os.path.join(tmpdirname, "video.mp4")79        download_yt_audio(yt_url, filepath)80        with open(filepath, "rb") as f:81            inputs = f.read()82 83    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)84    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}85 86    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]87 88    return html_embed_str, text89 90 91demo = gr.Blocks()92 93mf_transcribe = gr.Interface(94    fn=transcribe,95    inputs=[96        gr.inputs.Audio(source="microphone", type="filepath", optional=True),97        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),98    ],99    outputs="text",100    layout="horizontal",101    theme="huggingface",102    title="Whisper Large V3: Transcribe Audio",103    description=(104        "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the OpenAI Whisper"105        f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"106        " of arbitrary length."107    ),108    allow_flagging="never",109)110 111file_transcribe = gr.Interface(112    fn=transcribe,113    inputs=[114        gr.inputs.Audio(source="upload", type="filepath", optional=True, label="Audio file"),115        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),116    ],117    outputs="text",118    layout="horizontal",119    theme="huggingface",120    title="Whisper Large V3: Transcribe Audio",121    description=(122        "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the OpenAI Whisper"123        f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"124        " of arbitrary length."125    ),126    allow_flagging="never",127)128 129yt_transcribe = gr.Interface(130    fn=yt_transcribe,131    inputs=[132        gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),133        gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe")134    ],135    outputs=["html", "text"],136    layout="horizontal",137    theme="huggingface",138    title="Whisper Large V3: Transcribe YouTube",139    description=(140        "Transcribe long-form YouTube videos with the click of a button! Demo uses the OpenAI Whisper checkpoint"141        f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe video files of"142        " arbitrary length."143    ),144    allow_flagging="never",145)146 147with demo:148    gr.TabbedInterface([mf_transcribe, file_transcribe, yt_transcribe], ["Microphone", "Audio file", "YouTube"])149 150demo.launch(enable_queue=True)151 152