Kaldi-Installation

1. Install Kaldi

   • In desired installation folder (e.g. /opt)

     git clone https://github.com/kaldi-asr/kaldi.git kaldi --origin upstream
     cd kaldi

   • follow install instructions in INSTALL:
     1. go to tools/ and follow INSTALL instructions there.

        extras/check_dependencies.sh
        make

     2. go to src/ and follow INSTALL instructions there.

        ./configure --shared
        make depend -j 4
        make -j 4

        (if this fails (out of memory) run without "-j 4" option or with "-j 2" instead, or give the vm more memory)

2. Install some dependencies

   sudo apt-get install build-essential pkg-config python3-pip python3-dev python3-setuptools python3-wheel ffmpeg sox libatlas-base-dev libatlas3-base portaudio19-dev libasound-dev

3. Fix Library Paths

   • pkg-config
     • add a file called kaldi-asr.pc to /usr/local/lib/pkgconfig with content :

       kaldi_root=/opt/kaldi
       Name: kaldi-asr
       Description: kaldi-asr speech recognition toolkit
       Version: 5.5
       Requires: lapack-atlas blas-atlas
       Libs: -L${kaldi_root}/tools/openfst/lib -L${kaldi_root}/src/lib -lkaldi-decoder -lkaldi-lat -lkaldi-fstext -lkaldi-hmm -lkaldi-feat -lkaldi-transform -lkaldi-gmm -lkaldi-tree -lkaldi-util -lkaldi-matrix -lkaldi-base -lkaldi-nnet3 -lkaldi-online2 -lkaldi-cudamatrix -lkaldi-ivector -lfst
       Cflags: -I${kaldi_root}/src -I${kaldi_root}/tools/openfst/include

       (change kaldi_root to wherever it is installed in step1)
   • ld
     • add a file to /etc/ld.so.conf.d called kaldi.conf with a line containing the folders where kaldi tools and kaldi src libraries are stored, eg: /opt/kaldi/src/lib /opt/kaldi/tools/openfst/lib
     • if directory doesn't exist, instead add these line to /etc/ld.so.conf
     • then run

       sudo ldconfig

4. Create a symbolic link because one of the pip packages expect atlas in this location: ln -s /usr/include/x86_64-linux-gnu/atlas /usr/include/atlas

5. Install python dependencies (best under a python3 virtual env):

pip install wheel numpy nltk cython
pip install py-kaldi-asr py-nltools

6. If you encounter an error about the atlas package, do this:

   • create a file named atlas.pc under /usr/lib/pkgconfig -write following data inside(paths may differ on your system - check if the path exists):

   	Name: atlas
   	Version: 3.10.3
   	Description: Debug information for package atlas
   	Cflags: -I/usr/include/x86_64-linux-gnu/atlas
   	Libs: -L/usr/include/x86_64-linux-gnu/atlas

Running examples:

1. Decode microphone input in real time:

python3 kaldi_decode_live.py -m PATH_TO_MODEL

2. Decode a wav file:

python3 kaldi_decode_wav.py -m PATH_TO_MODEL input.wav

Change input.wav to whichever file you want to decode. If the file is not in the current folder, provide a full path to the file. THe wav file must be single chanel 16 bit.

Further training

TO adapt the model or train another one refer to Zamia-speech project.https://github.com/gooofy/zamia-speech It is much easier than using plain Kaldi but it is still not beginner friendly.