International Symposium on Chinese Spoken Language Processing (ISCSLP 2002)

Taipei, Taiwan
August 23-24, 2002

Fast Likelihood Computation Method Using Block-Diagonal Covariance Matrices in Hidden Markov Model

Rui Wang, Xuan Zhu, Yining Chen, Jia Liu, Runsheng Liu

Department of Electronics Engineering, Tsinghua University, Beijing, China

The paper presented a novel method to speed up the likelihood computation of the speech recognition system based continuous Hidden Markov Model (CHMM). The block-diagonal covariance matrices were applied in the method and the technique to construct an optimal block-diagonal matrix was introduced. The experimental results demonstrated that the block-diagonal covariance matrices could achieve a large improvement in recognition speed without significant decrease of recognition rate compared with baseline system.

Full Paper

Bibliographic reference.  Wang, Rui / Zhu, Xuan / Chen, Yining / Liu, Jia / Liu, Runsheng (2002): "Fast likelihood computation method using block-diagonal covariance matrices in hidden Markov model", In ISCSLP 2002, paper 118.