A new algorithm to reduce the amount of calculation in the likelihood computation of continuous mixture HMM(CMHMM) with block-diagonal covariance matrices while retaining high recognition rate is proposed. The block matrices are optimized by minimizing difference between the output probability calculated with full covariance matrices and that calculated with block-diagonal covariance matrices. The idea was implemented and tested on a continuous number recognition task.
Cite as: Koshiba, R., Tachimori, M., Kanazawa, H. (1998) A flexible method of creating HMM using block-diagonalization of covariance matrices. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0197, doi: 10.21437/ICSLP.1998-182
@inproceedings{koshiba98_icslp,
author={Ryosuke Koshiba and Mitsuyoshi Tachimori and Hiroshi Kanazawa},
title={{A flexible method of creating HMM using block-diagonalization of covariance matrices}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0197},
doi={10.21437/ICSLP.1998-182}
}