Unsupervised training for acoustic models of speech

  • Gregor Donaj University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Mathematics, Maribor, Slovenia. https://orcid.org/0000-0002-0297-2714
  • Andrej Žgank University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Maribor, Slovenia.
  • Mirjam Sepesy Maučec University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Maribor, Slovenia. https://orcid.org/0000-0003-0215-513X
DOI: https://doi.org/10.18690/analipazu.3.2.69-74.2013
Keywords: acoustical models, speech recognition, unsupervised training

Abstract

This paper presents unsupervised acoustical model training for automatic speech recognition. The main advantage of this training method is its speed and cost effectiveness compared to the manual transcription of speech, which is needed for supervised training. We present two methods of unsupervised training and test them on a large vocabulary continuous speech recognition system in the Broadcast News domain.

Downloads

Download data is not yet available.

Author Biographies

Gregor Donaj, University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Mathematics, Maribor, Slovenia.

Maribor, Slovenija. E-mail: gregor.donaj@um.si

Andrej Žgank, University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Maribor, Slovenia.

Maribor, Slovenia. E-mail: andrej.zgank@um.si

Mirjam Sepesy Maučec, University of Maribor, Faculty of Electrical Engineering, Computer Science and Informatics, Maribor, Slovenia.

Maribor, Slovenia. E-mail: mirjam.sepesy@um.si

Published
2022-06-06
How to Cite
Donaj G., Žgank A., & Sepesy Maučec M. (2022). Unsupervised training for acoustic models of speech. Anali PAZU, 3(2), 69-74. https://doi.org/10.18690/analipazu.3.2.69-74.2013
Issue
Vol 3 No 2 (2013): ANALI PAZU
Section
Prispevki