About

This package includes files for building Arabic ASR using the GALE database from LDC and the Kaldi Speech Recognition Toolkit. The test set is mix of conversational and report speech.

Related publications

  • Ali, A.; Zhang, Y.; Cardinal, P.; Dahak, N.; Vogel, S.; Glass, J., A complete KALDI recipe for building Arabic speech recognition systems, Spoken Language Technology Workshop (SLT), pp.525,529, 7-10 Dec. 2014. [BibTeX]
    @INPROCEEDINGS{aaliArabicKaldi, author={Ali, A. and Yifan Zhang and Cardinal, P. and Dahak, N. and Vogel, S. and Glass, J.}, booktitle={Spoken Language Technology Workshop (SLT), 2014 IEEE}, title={A complete KALDI recipe for building Arabic speech recognition systems}, year={2014}, month={Dec}, pages={525-529}, doi={10.1109/SLT.2014.7078629} }

Download

Installation

  1. Install and compile Kaldi: http://kaldi.sourceforge.net/
  2. Untar the gale_recipe.tar into egs folder: tar xvf gale_recipe.tar -C kaldi-trunk/egs
  3. Modify run.sh:
    1. Adjust the number of jobs accordingly, default nJobs=120 It will also depend if you use queue or local machine, look at cmd.sh
    2. Change the data settings to point to GALE database: example: LDC2013S02_1=/alt/data/speech/LDC/LDC2013S02/gale_p2_arb_bc_speech_p1_d1
  4. Start run.sh

The script will build GMM, GMM+MPE, GMM+bMMI, SGMM+fMLLR, and SGMM+bMMI The RESULTS file has the WER if you use the script along with QCRI pronunciation dictionary.