mau_local_stt

A Maubot to transcribe audio messages in matrix rooms using local open-source libraries

Installation

1. FFmpeg must be in $PATH
2. Activate the maubot virtual environment (source ./bin/activate), and run
   • pip install whispercpp numpy - if you want to use whisper as the backend.
   • pip install vosk - if you want to use vosk as the backend.
3. Download maulocalstt from the releases (or download the repository and build with mbc build), and upload it to maubot.
4. Download a model for your backend:
   • For wisper, download a model from https://huggingface.co/ggerganov/whisper.cpp and place it under models/whispercpp
   • For vosk, download a zipped model from https://alphacephei.com/vosk/models and unpack it into models/vosk
5. Create an instance of the bot, and update the configuration:
   • For whisper, specify
     • model_name - the name of the model you downloaded (the name of the file without the ggml- and .bin)
     • language - the language the audio will be in (you can set it to auto for whisper to auto-detect the language)
     • translate - if you want wisper to translate the transcription to english (true or false)
   • For vosk, specify
     • model_path - the path to the top directory of the model you downloaded (the one with the folders am conf graph etc.), either absolute or related to maubot's working directory.

Usage

Simply invite the bot to a room, and it will reply to all audio messages with their transcription