Current time-synchronous beam-search algorithm is improved from two aspects for speeding up large vocabulary continuous speech recognition. Single-triphone-tree structure is proposed to take instead of the tree copy technique for simplifying the search computation and saving the memory . By one kind of special-designed token propagation strategy, the n-gram language model can be integrated into the single-tree search algorithm. Moreover, a lexical tree based language model format is defined to store the pre-computed lookahead probabilities by deploying the back-off mechanism to limit the memory requirement within a manageable range, and in this way the online computation of lookahead language model can be effectively accelerated. Finally a language-independent general decoder is implemented, including English WSJ20k and Mandarin51k dictation system. Experiment results indicates that high accuracy recognition result can be attained only in the first pass by the single-triphone tree search algorithm, and search efforts can be reduced by 16% with the pre-computing lookahead LM technique.
Cite as: Zhao, Q., Lin, Z., Yuan, B., Yan, Y. (2000) Improvements in search algorithm for large vocabulary continuous speech recognition. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 4, 306-309, doi: 10.21437/ICSLP.2000-812
@inproceedings{zhao00f_icslp,
author={Qingwei Zhao and Zhiwei Lin and Baosheng Yuan and Yonghong Yan},
title={{Improvements in search algorithm for large vocabulary continuous speech recognition}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 4, 306-309},
doi={10.21437/ICSLP.2000-812}
}