This paper presents a method to integrate the class bigram language model effectively to the lexical tree decoder. The method reduces the memory requirement and search effort in comparison with the conventional lexical tree search with word bigram language model. The decoder is based on a time-synchronous beam search, using cross-word triphone acoustic model. To demonstrate its effectiveness, the algorithm is tested with a stock query task. Experimental results show that the lexical tree decoder based on a class bigram can reduce the search space by 11.8%. By using class-bigram look-ahead, the memory cost for storing the look-ahead probability can also achieve a saving of 73% without degradation in recognition accuracy.
Cite as: Choi, W.N., Wong, Y.W., Lee, T., Ching, P.C. (2000) Lexical tree decoding with a class-based language model for Chinese speech recognition. Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000), vol. 1, 174-177, doi: 10.21437/ICSLP.2000-43
@inproceedings{choi00_icslp,
author={W. N. Choi and Y. W. Wong and Tan Lee and P. C. Ching},
title={{Lexical tree decoding with a class-based language model for Chinese speech recognition}},
year=2000,
booktitle={Proc. 6th International Conference on Spoken Language Processing (ICSLP 2000)},
pages={vol. 1, 174-177},
doi={10.21437/ICSLP.2000-43}
}