Multi-ASR Toolkit is a flexible and extensible speech recognition toolkit supporting multiple backend engines such as Whisper, Faster-Whisper, WhisperX, SpeechRecognition, and Vosk. It provides both a command-line interface and a web-based interface via Gradio, facilitating easy transcription of audio files using various ASR models.

• Python 3.10 or higher
• Gradio requires Python 3.10 or higher.
• pygame
• pydub
• ffmpeg: for convert mp3 to wav
• PyTorch 2.1+, TensorFlow 2.6+
• transformers
• SpeechRecognition
• Whisper
• Faster Whisper
• MLX Whisper
• demucs
• yt-dlp

1. Python packages

   $ pip3 install -r requirements.txt

2. ffmpeg

   # Ubuntu
   $ sudo apt install ffmpeg
   
   # Mac
   $ brew install ffmpeg

   For Windows, you can refer to this website: ffmpeg install

# python app.py --mode cli <wav/mp3 file> 
$ python app.py --mode cli data/test.mp3 

# python app.py --mode cli <wav/mp3 file> --backend <asr backend> --language <language> --model-size <model size>
$ python app.py --mode cli data/test.mp3 --backend faster-whisper --language en --model-size base

$ python3 app.py

• Open a new incognito/private window and log in to your YouTube account.

• In the same tab, open https://www.youtube.com/robots.txt, ensuring that only this tab is using the login session.

• Use a browser extension (e.g., "cookies.txt" for Chrome) to export the youtube.com cookies for this session to a file named cookies.txt, and then immediately close the incognito window.

• Using the manually exported cookies.txt in Python.

  Place your cookies.txt in a fixed path, and then specify it in ydl_opts like this:

  ydl_opts = {
  'format': 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/mp4',
  'outtmpl': filepath,
  'verbose': True,          # 用于调试，正式可去掉
  'merge_output_format': 'mp4',
  'cookies': '<Path>/youtube_cookies.txt',
  }

• I just copied the demucs folder from the demucs repo into the backends folder.

• Demucs APIs docs

• abus-aikorea/kara-audio