Edward WILLEMS, Tetsuo KOSAKA, Shigeki SAGAYAMA
Variance Spreading for Better Recognition
using HM-Nets : a deterministic approach
Abstract:Empirical observation has shown that substantial improvements in speaker-independent
recognition using HM-Nets can be achieved, if all the variances in the HM-Net are spread by a
same factor. Optimal choice of the value of this factor can increase the recognition rate by as
much as 20 percent. However, the value of the factor is currently determined heuristically for
each speaker, which undermines the usefulness of this adaptation method. The purpose of the
study presented in this report was to investigate the underlying parameters in the speech signal
which determine the optimal spread factor value, in order to devise a more robust procedure
for speaker-adaptation. The study concentrated on finding a correlation between the optimal
spread factor value and the difference between the parameter distributions of the HM-Net,
and those determined directly from the speaker's speech data. The data from five out of six
speakers showed a correlation between Bhatacharyya distance, which is a measure of both
mean and variance difference, and optimal spread factor value.