Efficient reduction of Gaussian components using MDL criterion for HMM-based speech recognition

被引:0
作者
Shinoda, K [1 ]
Iso, K [1 ]
机构
[1] Univ Tokyo, Grad Sch Informat Sci & Technol, Bunkyo Ku, Tokyo 1138656, Japan
来源
2002 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOLS I-IV, PROCEEDINGS | 2002年
关键词
D O I
暂无
中图分类号
O42 [声学];
学科分类号
070206 ; 082403 ;
摘要
A method is proposed to reduce the number of Gaussian components in continuous density hidden Markov models (HMMs). As its initial model, the method employs a well-trained, large-sized HAM in which the components of each state's Gaussian mixture probability density function are clustered into a binary tree. For each state, a subset of Gaussian components is chosen from the Gaussian tree on the basis of the minimum description length (MDL) criterion. By varying the penalty coefficient for large size models in the MDL criterion, it is possible to obtain the total number of Gaussian components desired for smaller models. In our experimental evaluations, the proposed method successfully reduced the number of Gaussian components by 75%, with only 1% degradation in recognition accuracy.
引用
收藏
页码:869 / 872
页数:4
相关论文
共 11 条
[1]  
CHAN YC, 2000, P ICSLP2000
[2]  
Chen S., 1999, P EUR C SPEECH COMM, P1087
[3]  
FRITSCH J, 1999, P EUROSPEECH99, P1103
[4]  
HWANG MY, 1988, P ICASSP98, P669
[5]  
JIA Y, 2000, P ICSLP2000
[6]   UNIVERSAL CODING, INFORMATION, PREDICTION, AND ESTIMATION [J].
RISSANEN, J .
IEEE TRANSACTIONS ON INFORMATION THEORY, 1984, 30 (04) :629-636
[7]  
SANKAR A, 1999, P EUROSPEECH99, P1711
[8]  
Shinoda K., 2000, Journal of the Acoustical Society of Japan (E), V21, P79, DOI 10.1250/ast.21.79
[9]  
SHINODA K, 2001, THESIS TOKYO I TECHN
[10]  
SHINODA K, 1997, P EUROSPEECH, V1, P99