Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Comparative Evaluation of Segmental Unit Input HMM And Conditional Density HMM

Kazumasa Yamamoto, Seiichi Nakagawa

Toyohashi University of Technology, Department of Information and Computer Sciences, Toyohashi, Japan

The standard HMM cannot express the time variant features during staying at the same state. We tried to capture the dynamic changes by using segmental statistics. We propose a new speech recognition method by the combination of HMM and segmental statistics. Using segmental statistics, since the dimension of parameters increases, it results in a lesser precision in the estimation of covariance matrix. Therefore we used methods for compressing dimension and reducing computation, such as K-L expansion and Modified Quadratic Discriminant Function(MQDF). This method outperformed traditional methods such as a conditional density HMM with the correlation between two frames and an HMM using regression coefficients as dynamic features.

Full Paper

Bibliographic reference.  Yamamoto, Kazumasa / Nakagawa, Seiichi (1995): "Comparative evaluation of segmental unit input HMM and conditional density HMM", In EUROSPEECH-1995, 1615-1618.