app.py

import torch
import time

import gradio as gr
import yt_dlp as youtube_dl
from transformers import pipeline
from transformers.pipelines.audio_utils import ffmpeg_read

import tempfile
import os

BATCH_SIZE = 8
FILE_LIMIT_MB = 1000
YT_LENGTH_LIMIT_S = 3600  # limit to 1 hour YouTube files

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


def transcribe(model, audio, task):
    if audio is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")

    pipe = pipeline(
        task="automatic-speech-recognition",
        model=model,
        chunk_length_s=30,
        device=device,
    )
    text = pipe(audio, batch_size=BATCH_SIZE, generate_kwargs={"language": "latvian", "task": task}, return_timestamps=True)["text"]
    return 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 download_yt_audio(yt_url, filename):
    info_loader = youtube_dl.YoutubeDL()
    
    try:
        info = info_loader.extract_info(yt_url, download=False)
    except youtube_dl.utils.DownloadError as err:
        raise gr.Error(str(err))
    
    file_length = info["duration_string"]
    file_h_m_s = file_length.split(":")
    file_h_m_s = [int(sub_length) for sub_length in file_h_m_s]
    
    if len(file_h_m_s) == 1:
        file_h_m_s.insert(0, 0)
    if len(file_h_m_s) == 2:
        file_h_m_s.insert(0, 0)
    file_length_s = file_h_m_s[0] * 3600 + file_h_m_s[1] * 60 + file_h_m_s[2]
    
    if file_length_s > YT_LENGTH_LIMIT_S:
        yt_length_limit_hms = time.strftime("%HH:%MM:%SS", time.gmtime(YT_LENGTH_LIMIT_S))
        file_length_hms = time.strftime("%HH:%MM:%SS", time.gmtime(file_length_s))
        raise gr.Error(f"Maximum YouTube length is {yt_length_limit_hms}, got {file_length_hms} YouTube video.")
    
    ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
    
    with youtube_dl.YoutubeDL(ydl_opts) as ydl:
        try:
            ydl.download([yt_url])
        except youtube_dl.utils.ExtractorError as err:
            raise gr.Error(str(err))


def yt_transcribe(model, 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()

    pipe = pipeline(
        task="automatic-speech-recognition",
        model=model,
        chunk_length_s=30,
        device=device,
    )
    inputs = ffmpeg_read(inputs, pipe.feature_extractor.sampling_rate)
    inputs = {"array": inputs, "sampling_rate": pipe.feature_extractor.sampling_rate}

    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"language": "latvian", "task": task}, return_timestamps=True)["text"]

    return html_embed_str, text


demo = gr.Blocks()

transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        # gr.Markdown(
        #     """
        #     Test Latvian speech recognition (STT) models. Three models are available:
        #     * [tiny](https://huggingface.co/RaivisDejus/whisper-tiny-lv) - Fastest, requiring least RAM, but also least accurate
        #     * [small](https://huggingface.co/RaivisDejus/whisper-small-lv) - Reasonably fast, reasonably accurate, requiring reasonable amounts of RAM
        #     * [large](https://huggingface.co/AiLab-IMCS-UL/whisper-large-v3-lv-late-cv17) - Most accurate, developed by scientists from [ailab.lv](https://ailab.lv/). Requires most RAM and for best performance should be run on a GPU.
        #     """
        # ),
        gr.Dropdown([
            ("tiny", "RaivisDejus/whisper-tiny-lv"),
            ("small", "RaivisDejus/whisper-small-lv"),
            ("large", "AiLab-IMCS-UL/whisper-large-v3-lv-late-cv17")
        ], label="Model", value="RaivisDejus/whisper-small-lv"),
        gr.Audio(sources=["upload", "microphone"],type="filepath", label="Audio"),
        gr.Radio([("Transcribe", "transcribe"), ("Translate to English", "translate",)], label="Task", value="transcribe"),
    ],
    outputs=gr.Textbox(label="Transcription", lines=10),
    title="Latvian speech recognition: Transcribe Audio",
    description=("""
        Test Latvian speech recognition (STT) models. Three models are available:
        
        * [tiny](https://huggingface.co/RaivisDejus/whisper-tiny-lv) - Fastest, requiring least RAM, but also least accurate
        
        * [small](https://huggingface.co/RaivisDejus/whisper-small-lv) - Reasonably fast, reasonably accurate, requiring reasonable amounts of RAM
        
        * [large](https://huggingface.co/AiLab-IMCS-UL/whisper-large-v3-lv-late-cv17) - Most accurate, developed by scientists from [ailab.lv](https://ailab.lv/). Requires most RAM and for best performance should be run on a GPU.
        """
    ),
    allow_flagging="never",
)

yt_transcribe = gr.Interface(
    fn=yt_transcribe,
    inputs=[
        gr.Dropdown([
            ("tiny", "RaivisDejus/whisper-tiny-lv"),
            ("small", "RaivisDejus/whisper-small-lv"),
            ("large", "AiLab-IMCS-UL/whisper-large-v3-lv-late-cv17")
        ], label="Model", value="RaivisDejus/whisper-small-lv"),
        gr.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
        gr.Radio([("Transcribe", "transcribe"), ("Translate to English", "translate",)], label="Task", value="transcribe")
    ],
    outputs=["html", "text"],
    title="Latvian speech recognition: Transcribe YouTube",
    description=(
        "Transcribe long-form YouTube videos with the click of a button! Demo uses the checkpoint"
    ),
    allow_flagging="never",
)

with demo:
    gr.TabbedInterface([transcribe, yt_transcribe], ["Microphone / Audio file", "YouTube"])

demo.queue(max_size=10)
demo.launch()