Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Discriminative Training of HMM Based Speech Recognizer with Gradient Projection Method

Qiang Huo (1,2), Chorkin Chan (1)

(1) Department of Computer Science, The University of Hong Kong, Hong Kong
(2) ATR Interpreting Telecommunications Research Labs., Seika-cho Soraku-gun, Kyoto, Japan

In this paper, in order to examine the viability and characteristics of our previously proposed gradient projection method for HMM training, the method is applied to discriminatively train the DHMM parameters with the objective of minimizing the recognition error rate and a series of experiments have been conducted. The experiments involve speaker independent recognition of the highly confusable E-set of the English alphabet. Three kinds of training schemes, namely, batch training, sequential block training and corrective training sue studied. The experimental results show that the minimum recognition error objective function is a viable formulation that can be optimized by the gradient projection method.

Full Paper

Bibliographic reference.  Huo, Qiang / Chan, Chorkin (1995): "Discriminative training of HMM based speech recognizer with gradient projection method", In EUROSPEECH-1995, 101-104.