MPhil Thesis Defence


"Kernel Eigenspace-based MLLR Adaptation"

By

Mr. Wend-Huu Hsiao


Abstract

Kernel methods have been applied to improve the performance of existing
eigenvoice-based adaptation methods. Several adaptation methods including
kernel eigenvoice adaptation (KEV) and embedded kernel eigenvoice
adaptation (eKEV) report promising results. The basic idea of these
kernel-based adaptation methods is to exploit possible nonlinearity among
different speakers.

In this thesis, a variant of eigenspace-based maximum likelihood linear
regression (EMLLR) adaptation, named kernel eigenspace-based MLLR
adaptation (KEMLLR), is proposed. It adopts kernel methods to map the MLLR
transformation matrices to a kernel-induced high dimensional space and
kernel principal component analysis (KPCA) is used to derive a set of
kernel eigenmatrices in the feature space. A new speaker is then
represented by a linear combination of the leading kernel eigenmatrices in
the feature space, and the eigenmatrix weights are optimized by using BFGS
quasi-Newton algorithm.

In the Resource Management adaptation task using 10-mixture monophone
hidden Markov models, KEMLLR gives encouraging results in rapid speaker
adaptation. It is found that when only 5s of adaptation speech are
available, EMLLR successfully reduces the word error rate (WER) by 7.82%,
and KEMLLR can reduce the WER by 11.4%. When 10s of adaptation speech are
provided, MLLR using full transformation becomes effective and matches the
performance of KEMLLR. EMLLR, eKEV and MLLR using diagonal transformation
do not perform as well as KEMLLR in the experiments.

The time complexities of various adaptation methods including EV, MLLR,
EMLLR, KEV, eKEV, and KEMLLR are also analyzed under some mild
assumptions. Comparisons are made and the analysis helps understanding the
efficiency of various methods.


Date:				Friday, 26 November 2004

Time:				2:00p.m.-4:00p.m.

Venue:				Room 2407
				Lifts 17-18

Committee Members:		Prof. Brian Mak (Supervisor)
				Prof. James Kwok (Chairperson)
				Prof. Tan Lee (CUHK)



**** ALL are Welcome ****