ISCA Archive ICSLP 1990
ISCA Archive ICSLP 1990

Speaker adaptation of continuous parameter HMM

Yoshimitsu Hirata, Seiichi Nakagawa

As an speaker adaptation method of continuous parameter HMM, we adapted mean vectors which are a part of parameters of multi-dimensional normal distributions. We regard a set of mean vectors belonging to each HMM as a codebook. The unsupervised adaptation algorithm modifies the mean vectors by using vector-quantization error-vectors for a test speaker. For two persons, 23 Japanese phoneme recognition accuracy was improved from 62% of non-adaptation into 73% after unsupervised adaptation and into 82% after supervised adaptation, respectively. Also, we describe adaptation results through multi-speaker mode HMM and comparison between the improvement by speaker adaptation and the degradation by vector-quantization of input vectors on speaker dependent mode HMM.