app.py

import torch
import spaces
import gradio as gr
from pytube import YouTube
from transformers import pipeline, WhisperProcessor, WhisperForConditionalGeneration
from transformers.pipelines.audio_utils import ffmpeg_read

import tempfile
import os

MODEL_NAME = "MohamedRashad/Arabic-Whisper-CodeSwitching-Edition"
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000*3
YT_LENGTH_LIMIT_S = 60*60*3  # limit to 3 hour YouTube files

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

processor = WhisperProcessor.from_pretrained(MODEL_NAME)
model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)

pipe = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=30,
    device=device,
)

@spaces.GPU(duration=120)
def transcribe(inputs, task):
    if inputs is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")

    generate_kwargs = {"task": task, "language": "arabic" if task == "transcribe" else "english"}
    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs=generate_kwargs, return_timestamps=True)["text"]
    return  text


def _return_yt_html_embed(yt_url):
    video_id = YouTube(yt_url).video_id
    HTML_str = (
        f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
        " </center>"
    )
    return HTML_str

def download_yt_audio(yt_url, filename):
    yt = YouTube(yt_url)
    
    if yt.length > YT_LENGTH_LIMIT_S:
        raise gr.Error("YouTube video is too long! Please upload a video that is less than 1 hour long.")
    
    stream = yt.streams.filter(only_audio=True).first()
    stream.download(filename=filename)

def seconds_to_timestamp(seconds):
    total_seconds = int(seconds)
    hours = total_seconds // 3600
    minutes = (total_seconds % 3600) // 60
    remaining_seconds = seconds % 60
    return f"{hours:02d}:{minutes:02d}:{remaining_seconds:06.3f}"


def chunks_to_subtitle(chunks):
    subtitle = ""
    for chunk in chunks:
        start = seconds_to_timestamp(chunk["timestamp"][0])
        end = seconds_to_timestamp(chunk["timestamp"][1])
        text = chunk["text"]
        subtitle += f"{start} --> {end}\n{text}\n\n"
    return subtitle

@spaces.GPU(duration=120)
def yt_transcribe(yt_url, task):
    html_embed_str = _return_yt_html_embed(yt_url)

    with tempfile.TemporaryDirectory() as tmpdirname:
        filepath = os.path.join(tmpdirname, "video.mp4")
        download_yt_audio(yt_url, filepath)
        with open(filepath, "rb") as f:
            inputs = f.read()

    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}

    generate_kwargs = {"task": task, "language": "arabic" if task == "transcribe" else "english"}
    output = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs=generate_kwargs, return_timestamps=True)
    subtitle = chunks_to_subtitle(output["chunks"])

    return html_embed_str, subtitle


demo = gr.Blocks()

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="microphone", type="filepath"),
        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
    ],
    outputs="text",
    title="Arabic Whisper Code-Switching Edition: Transcribe Microphone",
    description=(
        "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the"
        f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"
        " of arbitrary length."
    ),
    allow_flagging="never",
)

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.Audio(sources="upload", type="filepath", label="Audio file"),
        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
    ],
    outputs="text",
    title="Arabic Whisper Code-Switching Edition: Transcribe Audio",
    description=(
        "Transcribe long-form microphone or audio inputs with the click of a button! Demo uses the"
        f" checkpoint [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe audio files"
        " of arbitrary length."
    ),
    allow_flagging="never",
)

yt_transcribe_demo = gr.Interface(
    fn=yt_transcribe,
    inputs=[
        gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
        gr.Radio(["transcribe", "translate"], label="Task", value="transcribe"),
    ],
    outputs=["html", "text"],
    title="Arabic Whisper Code-Switching Edition: Transcribe YouTube Video",
    description=(
        "Transcribe long-form YouTube videos with the click of a button! Demo uses the checkpoint"
        f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) and 🤗 Transformers to transcribe video files of"
        " arbitrary length."
    ),
    allow_flagging="never",
)

with demo:
    gr.TabbedInterface([mf_transcribe, file_transcribe, yt_transcribe_demo], ["Microphone", "Audio file", "YouTube"])

demo.queue().launch()