Training Discriminative HMM by Optimal Allocation of Gaussian Kernels

We propose to train Hidden Markov Model (HMM) by allocating Gaussian kernels non-uniformly across states so as to optimize a selected discriminative training criterion. The optimal kernel allocation problem is first formulated based upon a non-discriminative, Maximum Likelihood (ML) criterion and then generalized to incorporate discriminative ones. An effective kernel exchange algorithm is derived and tested on TIDIGITS, a speaker-independent (man, woman, boy and girl), connected digit recognition database. Relative 46-51% word error rate reductions are obtained comparing to the conventional uniformly allocated ML baseline. The recognition performance of discriminative kernel allocation is also consistently better than the non-discriminative ML based, nonuniform kernel allocation.

zhijie06.pdf
PDF file

In  Proc. of ISCSLP 2006

Publisher  Chinese and Oriental Language Information Processing Society
Copyright © 2007 by Chinese and Oriental Language Information Processing Society

Details

TypeProceedings
> Publications > Training Discriminative HMM by Optimal Allocation of Gaussian Kernels