Install Vosk

Vosk is a lightweight speech recognition (ASR) toolkit based on Kaldi that supports multiple languages and can run offline. The API is hosted at alphacep/vosk-api. We use it in our speech-related applications.

Install Dependencies

Vosk depends on Kaldi (see Install Kaldi).

Install Vosk (Ubuntu)

Assuming you have downloaded and built Kaldi in /opt/kaldi (or whatever directory of your choice; make sure this path contains the src/ directory), you can build Vosk as follows:

git clone https://github.com/alphacep/vosk-api
cd vosk-api/src
make -j KALDI_ROOT=/opt/kaldi

This will generate a libvosk.so file you can then link against in other libraries and applications.

Add GPU support

Assuming Kaldi has been built with GPU support, you can add it to Vosk by passing an additional flag to the make command. However, first inspect src/Makefile and remove -lnvToolsExt (like here).

make -j KALDI_ROOT=/opt/kaldi HAVE_CUDA=1

Download Vosk Models

All Vosk models can be downloaded from here.