First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

An Optimal Discriminative Training Method for Continuous Mixture Density HMMs

Shinobu Mizuta, Kunio Nakajima

Information Systems and Electronics Development Laboratory, Mitsubishi Electric Corporation, Kamakura, Japan

In this paper, we describe a training method for continuous mixture density HMM parameters, called optimal discriminative training. Conventional maximum likelihood estimation method for HMM training has a problem that the recognition performance is not considered in the training procedure. To solve the problem, a corrective training method has been already-proposed, but this method is applied to discrete HMMs, so the trained HMMs cannot avoid undesirable effects of VQ distortion. In this paper, the optimal discriminative training method (ODT) is described, appling the the basic concept of the corrective training to continuous mixture density HMMs, better recognition performance can be obtained as avoiding the VQ distortion. Prom word recognition experiments, we discuss the way to optimize each parameters of this training method, and by using the optimum value of parameters, we show the effectiveness of this method.

Full Paper

Bibliographic reference.  Mizuta, Shinobu / Nakajima, Kunio (1990): "An optimal discriminative training method for continuous mixture density HMMs", In ICSLP-1990, 245-248.