Fork of Kaldi for developing custom recognisers for Alex spoken dialogue system framework

News & info

• We use Docker so you can try easily our decoding demo

  • Run the demo using the two commands:

    1. download image docker pull ufaldsg/pykaldi

    2. run the demo docker run ufaldsg/pykaldi /bin/bash -c "cd online_demo; make gmm-latgen-faster; make online-recogniser; make pyonline-recogniser"

       • Note the demo downloads the pretrained models and test data which you may safe using docker commit functionality

  • Start exploring the demo source codes online_demo/pykaldi-online-latgen-recogniser.py and onl-rec/onl-rec-latgen-recogniser-demo.cc

  • Please note, that you need to change the source code of Pykaldi in the docker image to effect the demo behaviour when using docker.

• The Python wrapper of C++ OnlineLatticeRecogniser implements MFCC, LDA+MLLT, bMMI acoustic models since it was the best speaker independent setup.

• UPDATE: Since 11/18/2014 the Pykaldi fork uses the Kaldi official code (src/online2) which has very similar as our previous implementation (and was finished roughly 8 month after our implementations).

Install

• Our priority is to deploy it on Ubuntu 14.04 and also keep Travis running on Ubuntu 12.04
• Read INSTALL.rst and INSTALL first!
• INSTALL.rst contains instructions specific for this fork. INSTALL stores general instructions for Kaldi.

LICENSE

• This Kaldi fork is released under the Apache license, Version 2.0, which is also used by Kaldi itself.

• We also publicly released Czech and English data for kaldi/egs/vystadial_{cz,en} recipe under Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0) license:

  • Czech data https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4670-6
  • English data https://lindat.mff.cuni.cz/repository/xmlui/handle/11858/00-097C-0000-0023-4671-4

  Note that the data are automatically downloaded in the recipe scripts.

History

The fork presented three new Kaldi features in thesis of Ondrej Platek (see commit 8e534b16bb8a350): * Training scripts which can be used with standard Kaldi tools or with the new OnlineLatticeRecogniser.

The scripts for Czech and English support acoustic models obtained using MFCC, LDA+MLLT/delta+delta-delta feature transformations and acoustic models trained generatively or by MPE or bMMI training.

The new functionality was separated to different directories:

• pykaldi/src/onl-rec stores C++ code for OnlineLatticeRecogniser.
• pykaldi/pykaldi stores Python wrapper PyOnlineLatticeRecogniser.
• kaldi/egs/vystadial_{cz,en}/s5 stores training scripts. [merged to oficial Kaldi repo]
• kaldi/online_demo shows Kaldi standard decoder, OnlineLatticeRecogniser and PyOnlineLatticeRecogniser, which produce the exact same lattices using the same setup.

The OnlineLatticeRecogniser is used in Alex dialogue system (https://github.com/UFAL-DSG/alex).

In March 2014, the PyOnlineLatticeRecogniser recogniser was evaluated on Alex domain. See graphs evaluating OnlineLatticeRecogniser performance at http://nbviewer.ipython.org/github/oplatek/pykaldi-eval/blob/master/Pykaldi-evaluation.ipynb.

An example posterior word lattice output for one Czech utterance can be seen at http://oplatek.blogspot.it/2014/02/ipython-demo-pykaldi-decoders-on-short.html

Other info

• This Kaldi fork is developed under Vystadial project.
• Based on the Svn trunk of Kaldi project which is mirrored to branch svn-mirror.
• The svn trunk is mirrored via git svn. Checkout tutorials: Git svn, Svn branch in git