In this paper, we propose a rapid output probability calculation method in HMM based large vocabulary continuous speech recognition systems (LVCSRS). This method is based on time-skipping of calculation, clustering of probability density distributions, and pruning of calculation. Only distributions covering input feature vectors with high probabilities are used to calculate output probabilities strictly, and representative distributions for other distributions are used to calculate them approximately. Here a skipping method for likelihood calculation is adopted in the time domain. Using the rapid calculation method by clustering of probability density distributions, the recognition time in a LVCSRS system was reduced by about 40%. Using a pruning method of likelihood calculations on the way, it was further reduced by 25%. Finally, using time-skipping, the calculation time, furthermore, was reduced by 15% without compromising recognition accuracy.
Cite as: Nakagawa, S., Horibe, Y. (2001) A fast calculation method in LVCSRS by time-skipping and clustering of probability density distributions. Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001), 855-858, doi: 10.21437/Eurospeech.2001-135
@inproceedings{nakagawa01_eurospeech,
author={Seiichi Nakagawa and Yukihisa Horibe},
title={{A fast calculation method in LVCSRS by time-skipping and clustering of probability density distributions}},
year=2001,
booktitle={Proc. 7th European Conference on Speech Communication and Technology (Eurospeech 2001)},
pages={855--858},
doi={10.21437/Eurospeech.2001-135}
}