Sample models

Models for Kaldi. These models were trained on Polish Parliament data.

Included are:

• Phonetisaurus G2P model
• online chain TDNN acoustic model

Sample data

This package contains sample audio and text to test ASR. This data is from old sessions of the Polish Parliament.

Kaldi binaries

These are binaries for Kaldi that work with Google Colab.

This was generated from kaldi-asr/kaldi repository commit 070437fca7d3e09b405882ae1bfc373ae0f57ad1.

It also includes:

• openfst binaries
• phonetisaurus-g2pfst binary from Phonetisaurus G2P
• ngram and ngram-count binaries from SRILM

The binaries were created using nvidia-docker image nvidia/cuda:11.4.2-cudnn8-devel-ubuntu18.04.

This code to load it in Colab is as follows:

!wget https://github.com/danijel3/ASRforNLP/releases/download/v1.0/kaldi.tar.xz

!tar xvf kaldi.tar.xz -C / > /dev/null
%rm kaldi.tar.xz

!for f in $(find /opt/kaldi -name *.so*) ; do ln -sf $f /usr/local/lib/$(basename $f) ; done
!for f in $(find /opt/kaldi/src -not -name *.so* -type f -executable) ; do ln -s $f /usr/local/bin/$(basename $f) ; done
!for f in $(find /opt/kaldi/tools -not -name *.so* -type f -executable) ; do ln -s $f /usr/local/bin/$(basename $f) ; done

!ldconfig