Whispering

Streaming transcriber with whisper. Enough machine power is needed to transcribe in real time.

Notice

This repository has been archived. There are some alternatives.

• https://github.com/guillaumekln/faster-whisper
• https://github.com/ggerganov/whisper.cpp

Setup

pip install -U git+https://github.com/shirayu/[email protected]

# If you use GPU, install proper torch and torchaudio
# Check https://pytorch.org/get-started/locally/
# Example : torch for CUDA 11.6
pip install -U torch torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

If you get OSError: PortAudio library not found in Linux, install "PortAudio".

sudo apt -y install portaudio19-dev

Example of microphone

# Run in English
#  By the default, it needs to wait at least 30 seconds
whispering --language en --model tiny

• --help shows full options
• --model sets the model name to use. Larger models will be more accurate, but may not be able to transcribe in real time.
• --language sets the language to transcribe. The list of languages are shown with whispering -h
• --no-progress disables the progress message
• -t sets temperatures to decode. You can set several like -t 0.0 -t 0.1 -t 0.5, but too many temperatures exhaust decoding time
• --debug outputs logs for debug
• --vad sets VAD (Voice Activity Detection) threshold. The default is 0.5. 0 disables VAD and forces whisper to analyze non-voice activity sound period. Try --vad 0 if VAD prevents transcription.
• --output sets output file (Default: Standard output)
• --frame: the number of minimum frames of mel spectrogram input for Whisper (default: 3000. i.e. 30 seconds)

Parse interval

By default, whispering performs VAD for every 3.75 second. This interval is determined by the value of -n and its default is 20. When an interval is predicted as "silence", it will not be passed to whisper. If you want to disable VAD, please make VAD threshold 0 by adding --vad 0.

By default, whispering does not perform analysis until the total length of the segments determined by VAD to have speech exceeds 30 seconds. This is because the original Whisper assumes that the inputs are 30 seconds segments. However, if silence segments appear 16 times (the default value of --max_nospeech_skip) after speech is detected, the analysis is performed. You can make the length of segments smaller with --frame option (default: 3000), but it sacrifices accuracy because this is not expected input for Whisper.

Example of web socket

⚠ No security mechanism. Please make secure with your responsibility.

Run with --host and --port.

Host

whispering --language en --model tiny --host 0.0.0.0 --port 8000

Client

whispering --host ADDRESS_OF_HOST --port 8000 --mode client

You can set -n and other options.

For Developers

1. Install Python and Node.js

2. Install poetry to use poetry command

3. Clone and install libraries

   # Clone
   git clone https://github.com/shirayu/whispering.git
   
   # With poetry
   poetry config virtualenvs.in-project true
   poetry install --all-extras
   poetry run pip install -U torch torchaudio --extra-index-url https://download.pytorch.org/whl/cu116
   
   # With npm
   npm install

4. Run test and check that no errors occur

   poetry run make -j4

5. Make fancy updates

6. Make style

   poetry run make style

7. Run test again and check that no errors occur

   poetry run make -j4

8. Check typos by using typos. Just run typos command in the root directory.

   typos

9. Send Pull requests!

License

• MIT License
• Some codes are ported from the original whisper. Its license is also MIT License