Abstract
Our paper describes a two-pass recognition approach to the Czech speech recognition.We perform an automatic vocabulary adaptation after the first recognition pass, then we build a new language model using the adapted vocabulary and finally we run a second recognition pass. The aim of the vocabulary adaptation is to decrease the high OOV rate which is typical for the speech recognition of a highly inflectional languages such as Czech, Polish and other Slavic languages. Using the vocabulary adaptation algorithm we manage to decrease the OOV rate by 17% and, more importantly, our word accuracy improved by 11%. Therefore this approach can be successfully used in time uncritical recognition tasks.
This is a preview of subscription content, log in via an institution to check access.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Preview
Unable to display preview. Download preview PDF.
References
P. Geutner et al.: Transcribing Multilingual Broadcast News Using Hypothesis Driven Lexical Adaptation. DARPA Broadcast News Workshop. 1998
V. Radová, J. Psutka: UWB_S01 Corpus-A Czech Read-Speech Corpus. ICSLP 2000
A. Stolcke: SRILM-The SRI Language Modeling Toolkit. http://www.speech.sri.com/projects/srilm/
J. Hajič: Computational Morphology of Czech. Karolinum Press, Charles University. Prague. In prep.
S. Young et al.: The HTK Book. Entropic Inc. 1999
M. Mohri, M. Riley, F. C. N. Pereira: Weighted Finite-State Transducers in Speech Recognition. International Workshop on Automatic Speech Recognition: Challenges for the Next Millennium. 2000
Author information
Authors and Affiliations
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2001 Springer-Verlag Berlin Heidelberg
About this paper
Cite this paper
Ircing, P., Psutka, J. (2001). Two-Pass Recognition of Czech Speech Using Adaptive Vocabulary. In: Matoušek, V., Mautner, P., Mouček, R., Taušer, K. (eds) Text, Speech and Dialogue. TSD 2001. Lecture Notes in Computer Science(), vol 2166. Springer, Berlin, Heidelberg. https://doi.org/10.1007/3-540-44805-5_36
Download citation
DOI: https://doi.org/10.1007/3-540-44805-5_36
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-540-42557-1
Online ISBN: 978-3-540-44805-1
eBook Packages: Springer Book Archive