app.py

import os

import torch

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

MODEL_NAME = "NbAiLab/whisper-large-sme" #this always needs to stay in line 8 :D sorry for the hackiness
lang = "fi"

share = (os.environ.get("SHARE", "False")[0].lower() in "ty1") or None
auth_token = os.environ.get("AUTH_TOKEN") or True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

@spaces.GPU(duration=120)
def pipe(file):
    asr = pipeline(
        task="automatic-speech-recognition",
        model=MODEL_NAME,
        chunk_length_s=30,
        device=device,
        token=auth_token,
    )
    asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids(language=lang, task="transcribe")

def transcribe(file):
    text = pipe(file)["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 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"]

    return html_embed_str, text


demo = gr.Blocks()

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.components.Audio(sources=['upload', 'microphone'], type="filepath"),
    ],
    outputs="text",
    theme="huggingface",
    title="Whisper 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")],
    examples=[["https://www.youtube.com/watch?v=mukeSSa5GKo"]],
    outputs=["html", "text"],
    theme="huggingface",
    title="Whisper 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(share=True).queue()