Install Kaldi

Kaldi is a toolkit for speech recognition (ASR) written in C++. The official repository is kaldi-asr/kaldi, but we have forked it (roboticslab-uc3m/kaldi) to fix some compilation issues on top of alphacep/kaldi for use with the Vosk speech recognition toolkit by our speech-related applications.

Install Kaldi (Ubuntu)

These instructions have been adapted from the Vosk installation guide and its Dockerfile.

git clone https://github.com/roboticslab-uc3m/kaldi
cd kaldi-vosk-rl/tools
make -j openfst cub
extras/install_openblas_clapack.sh
cd ../src
./configure --shared --mathlib=OPENBLAS
sed -i 's: -O1 : -O3 :g' kaldi.mk
make -j online2 lm rnnlm

Add GPU support for Vosk

If available, the CUDA Toolkit is automatically detected and included in the configure step. After following the previous instructions, issue the following command in the src directory:

make -j cudafeat cudadecoder

You can now proceed with Install Vosk to complete the setup.