import time import os import re import base64 import torch import torchaudio import gradio as gr 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 import yt_dlp # Added import for yt-dlp MODEL_NAME = "NbAiLab/nb-whisper-large" lang = "no" with open("Logonew.png", "rb") as img_file: base64_image = base64.b64encode(img_file.read()).decode('utf-8') #logo_path = os.path.join(os.path.dirname(__file__), "Logo.png") max_audio_length= 30 * 60 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"Bruker enhet: {device}") @spaces.GPU(duration=60 * 2) def pipe(file, return_timestamps=False): asr = pipeline( task="automatic-speech-recognition", model=MODEL_NAME, chunk_length_s=28, device=device, token=auth_token, torch_dtype=torch.float16, model_kwargs={"attn_implementation": "flash_attention_2", "num_beams": 5} if FLASH_ATTENTION else {"attn_implementation": "sdpa", "num_beams": 5}, ) asr.model.config.forced_decoder_ids = asr.tokenizer.get_decoder_prompt_ids( language=lang, task="transcribe", no_timestamps=not return_timestamps, ) return asr(file, return_timestamps=return_timestamps, batch_size=24) def format_output(text): # Add a line break after ".", "!", ":", or "?" unless part of sequences like "..." #text = re.sub(r'(?', text) # Ensure line break after sequences like "..." or other punctuation patterns text = re.sub(r'(\.{3,}|[.!:?])', lambda m: m.group() + '
', text) return text def transcribe(file, return_timestamps=False): waveform, sample_rate = torchaudio.load(file) audio_duration = waveform.size(1) / sample_rate if audio_duration > max_audio_length: # Trim the waveform to the first 30 minutes waveform = waveform[:, :int(max_audio_length * sample_rate)] truncated_file = "truncated_audio.wav" torchaudio.save(truncated_file, waveform, sample_rate) file_to_transcribe = truncated_file truncated = True else: file_to_transcribe = file truncated = False if not return_timestamps: text = pipe(file_to_transcribe)["text"] formatted_text = format_output(text) else: chunks = pipe(file_to_transcribe, 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) formatted_text = "\n".join(text) if truncated: link="https://github.com/NbAiLab/nostram/blob/main/leverandorer.md" disclaimer = ( "\n\nDette er en demo. Det er ikke tillatt å bruke denne teksten i profesjonell sammenheng. " "Vi anbefaler at hvis du trenger å transkribere lengre opptak, så kjører du enten modellen lokalt " "eller sjekker denne siden for å se hvem som leverer løsninger basert på NB-Whisper: " "denne siden." ) formatted_text += f"

{disclaimer}" formatted_text += "

Transkribert med NB-Whisper demo" return formatted_text def _return_yt_html_embed(yt_url): video_id = yt_url.split("?v=")[-1] HTML_str = ( f'
' "
" ) return HTML_str def yt_transcribe(yt_url, return_timestamps=False): html_embed_str = _return_yt_html_embed(yt_url) ydl_opts = { 'format': 'bestaudio/best', 'outtmpl': 'audio.%(ext)s', 'postprocessors': [{ 'key': 'FFmpegExtractAudio', 'preferredcodec': 'mp3', 'preferredquality': '192', }], 'quiet': True, } with yt_dlp.YoutubeDL(ydl_opts) as ydl: ydl.download([yt_url]) text = transcribe("audio.mp3", return_timestamps=return_timestamps) return html_embed_str, text # Lag Gradio-appen uten faner demo = gr.Blocks() with demo: with gr.Column(): gr.HTML(f"") with gr.Column(scale=8): # Use Markdown for title and description gr.Markdown( """

NB-Whisper Demo

""" ) mf_transcribe = gr.Interface( fn=transcribe, inputs=[ gr.components.Audio(sources=['upload', 'microphone'], type="filepath"), gr.components.Checkbox(label="Inkluder tidsstempler"), ], outputs=gr.HTML(label="text"), #outputs="text", description=( "Transkriber lange lydopptak fra mikrofon eller lydfiler med et enkelt klikk! Demoen bruker den fintunede" f" modellen [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) og 🤗 Transformers til å transkribere lydfiler opp til 30 minutter." ), allow_flagging="never", #show_submit_button=False, ) # Uncomment to add the YouTube transcription interface if needed # yt_transcribe_interface = gr.Interface( # fn=yt_transcribe, # inputs=[ # gr.components.Textbox(lines=1, placeholder="Lim inn URL til en YouTube-video her", label="YouTube URL"), # gr.components.Checkbox(label="Inkluder tidsstempler"), # ], # examples=[["https://www.youtube.com/watch?v=mukeSSa5GKo"]], # outputs=["html", "text"], # title="Whisper Demo: Transkriber YouTube", # description=( # "Transkriber lange YouTube-videoer med et enkelt klikk! Demoen bruker den fintunede modellen:" # f" [{MODEL_NAME}](https://huggingface.co/{MODEL_NAME}) og 🤗 Transformers til å transkribere lydfiler av" # " vilkårlig lengde." # ), # allow_flagging="never", # ) # Start demoen uten faner demo.launch(share=share, show_api=False,allowed_paths=["Logo_2.png"]).queue()