Thesis: Automatic speech recognition using Kaldi

This repository contains the source code of the thesis and its presentation. If you want to read the thesis please read the most recent version text/tags/oplatek_thesis014.pdf.

(The most recent version should point to file with the highest number in the file name suffix. For older versions see the text/tags/ directory.)

I presented work overview and results of the thesis at the UFAL Monday seminar. The slides are available at ./presentation/tags/oplatek_ufal_seminar2014_04_14_v001.pdf. The slides for my defence in Czech can be downloaded at ./presentation/tags/oplatek__defence_cs.pdf.

Work described

The text of this thesis describes my work on projects listed below:

• Alex - Alex Dialogue System Framework(https://github.com/UFAL-DSG/alex) where I added following files and directories:

  • ASR component interfacing PyOnlineLatgenRecogniser - https://github.com/UFAL-DSG/alex/alex/components/asr/kaldi.py
  • Kaldi training scripts modified for Alex - https://github.com/UFAL-DSG/alex/alex/tools/kaldi/
  • Decoding graph (HCLG) scripts, and scripts for ASR evaluation - https://github.com/UFAL-DSG/alex/alex/PublicTransportInfoCs/hclg/

• The Kaldi toolkit - Speech recognition toolkit where I added following directories:

  • Implementation of OnlineLatgenRecogniser and utilities - https://github.com/UFAL-DSG/pykaldi/tree/master/src/onl-rec/
  • Python wrapper PyOnlineLatgenRecogniser and utilities - https://github.com/UFAL-DSG/pykaldi/tree/master/src/pykaldi/
  • Training scripts for acoustic modelling - https://github.com/UFAL-DSG/pykaldi/tree/master/egs/vystadial/s5/ The training scripts are also accepted in the Kaldi trunk but they are separated for Czech and English:
    • https://sourceforge.net/p/kaldi/code/HEAD/tree/trunk/egs/vystadial_en/
    • https://sourceforge.net/p/kaldi/code/HEAD/tree/trunk/egs/vystadial_cz/
  • Several demos using OnlineLatgenRecogniser and PyOnlineLatgenRecogniser - https://github.com/UFAL-DSG/pykaldi/tree/master/egs/vystadial/online_demo/

  I am developing the new features in https://github.com/UFAL-DSG/pykaldi repository which I mirror back to my Kaldi sandbox http://sourceforge.net/p/kaldi/code/HEAD/tree/sandbox/oplatek2 which can be merged to Kaldi trunk.

• pyfst - https://github.com/UFAL-DSG/pyfst - Python wrapper of OpenFST, where I improved installation and added several simple functions. Note I forked the original pyfst library.

• Pykaldi-eval - https://github.com/oplatek/pykaldi-eval - Repository for evaluation OnlineLatgenRecogniser written in IPython notebook. See interesting graphs.

• Alex documentation http://alex.readthedocs.org/en/latest/index.html

Note: The thesis is compiled for two side printing.