International Workshop on Hands-Free Speech Communication (HSC2001) April 9-11, 2001 Kyoto, Japan

Noise Compensation by a Sequential Kullback Proximal Algorithm

Kaisheng Yao (1), Kuldip K. Paliwal (1,2), Bertram E. Shi (3), and Satoshi Nakamura (1)

(1) ATR Spoken Language Translation Research Laboratories, Kyoto, Japan
(2) School of Microelectronics Engineering, Griffith University, Australia
(3) Rong Kong University of Science and Technology, Kowloon, Hong Kong, China

We present sequential parameter estimation in the framework of the Hidden Markov Models. The sequential algorithm is a sequential Kullback proximal algorithm, which chooses the Kullback-Liebler divergence as a penalty function for the maximum likelihood estimation. The scheme is implemented as filters. In contrast to algorithms based on the sequential EM algorithm, the algorithm has faster convergence rate in parameter estimation, and the computational complexity is proportional to the algorithms based on the sequential EM algorithm. In particular, we derive sequential noise parameter estimation for a model-based sequential noise compensation method for nonstationary noise environments. Noise parameter estimation, updating and speech recognition are carried out frame by frame. Simulation results have shown that the derived schemes can have faster convergence rate than the sequential noise compensation based on the sequential EM algorithm.

Full Paper

Bibliographic reference.  Yao, Kaisheng / Paliwal, Kuldip K. / Shi, Bertram E. / Nakamura, Satoshi (2001): "Noise compensation by a sequential Kullback proximal algorithm", In HSC2001, 139-142.