Pseudo-Swahili ASR for Kaldi

Uses almost no Swahili resources. Audio FSTs are trained from Voxforge's English.

Prerequisites

• The shell commands flac, gawk, swig, and wget.
  On Ubuntu, you might need to sudo apt install flac gawk swig wget.

• The Kaldi toolkit for automatic speech recognition.
  To install it, git clone https://www.github.com/kaldi-asr/kaldi.

• The SRI Language Modeling Toolkit.
  To add this to Kaldi, download the file srilm.tgz into kaldi/tools, and then (from kaldi/tools) ./install_srilm.sh.

• The Sequitur grapheme-to-phoneme converter.
  To add this to Kaldi, cd kaldi/tools && extras/install_sequitur.sh.
  (You might first need to sudo pip install numpy (for Python 2.7)).

Usage

Add these pseudo-Swahili scripts to Kaldi.

cd kaldi/egs
git clone https://www.github.com/uiuc-sst/pseudo-swahili
cd pseudo-swahili/s5
ln -s ../../wsj/s5/steps steps
ln -s ../../wsj/s5/utils utils

Get the Voxforge corpus of spoken English (this takes 45 minutes, and uses 25 GB of disk space).
./getdata.sh

Build the low-resource language model, vocabulary, etc. for Swahili.
cd pseudo-swahili/pseudo && ./a.sh

Build and test the speech recognizer.
cd pseudo-swahili/s5 && ./run.sh