화자적응을 위한 알고리즘 개선

Abstract

This paper studies the performances of MAP(Maximum a Posterior re-estimation), MLLR(Maximal Likelihood Linear Regression) and the combined method of each basic schemes. RCT(Regression Class Tree) is extended to MLLR. Database of korean 4 continuous digit numbers is used. Trained data are 4410 of man's voice and 3610 woman's one. We use test data of 980 of man's voice and 910 of woman's one. As a result, we obtain better recognition rate for system that combines each basic schemes than one that has a basic scheme. Also we get better recognition rate as class node, mixture number and the suitable tree depth increase.