app.py

import time
import os

import torch

import gradio as gr
import pytube as pt
import spaces
from transformers import AutoFeatureExtractor, AutoTokenizer, WhisperForConditionalGeneration, WhisperProcessor, pipeline
from huggingface_hub import model_info
try:
    import flash_attn
    FLASH_ATTENTION = True
except ImportError:
    FLASH_ATTENTION = False


MODEL_NAME = "NbAiLab/nb-whisper-large"
lang = "no"

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=60 * 2)
def pipe(file, return_timestamps=False):
    # model = WhisperForConditionalGeneration.from_pretrained(MODEL_NAME, torch_dtype=torch.float16, low_cpu_mem_usage=True)
    # model.to(device)
    # processor = WhisperProcessor.from_pretrained(MODEL_NAME)
    # model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
    # model.generation_config.cache_implementation = "static"
    asr = pipeline(
        task="automatic-speech-recognition",
        model=MODEL_NAME,
        # tokenizer=AutoTokenizer.from_pretrained(MODEL_NAME),
        # feature_extractor=AutoFeatureExtractor.from_pretrained(MODEL_NAME),
        chunk_length_s=30,
        device=device,
        token=auth_token,
        torch_dtype=torch.float16,
        model_kwargs={"attn_implementation": "flash_attention_2"} if FLASH_ATTENTION else {"attn_implementation": "sdpa"},
    )
    asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids(
        language=lang,
        task="transcribe",
        no_timestamps=not return_timestamps,
    )
    # asr.model.config.no_timestamps_token_id = asr.tokenizer.encode("<|notimestamps|>", add_special_tokens=False)[0]
    return asr(file, return_timestamps=return_timestamps, batch_size=24)

def transcribe(file, return_timestamps=False):
    if not return_timestamps:
        text = pipe(file)["text"]
    else:
        chunks = pipe(file, return_timestamps=True)["chunks"]
        text = []
        for chunk in chunks:
            start_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][0])) if chunk["timestamp"][0] is not None else "??:??:??"
            end_time = time.strftime('%H:%M:%S', time.gmtime(chunk["timestamp"][1])) if chunk["timestamp"][1] is not None else "??:??:??"
            line = f"[{start_time} -> {end_time}] {chunk['text']}"
            text.append(line)
        text = "\n".join(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, return_timestamps=False):
    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 = transcribe("audio.mp3", return_timestamps=return_timestamps)

    return html_embed_str, text


demo = gr.Blocks()

mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[
        gr.components.Audio(sources=['upload', 'microphone'], type="filepath"),
        gr.components.Checkbox(label="Return timestamps"),
    ],
    outputs="text",
    title="NB-Whisper Demo",
    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"),
        gr.components.Checkbox(label="Return timestamps"),
    ],
    examples=[["https://www.youtube.com/watch?v=mukeSSa5GKo"]],
    outputs=["html", "text"],
    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
    ], [
        "Transkriber Lyd",
        "Transkriber YouTube"
    ])

demo.launch(share=share).queue()