UBM Based Speaker Selection and Model Re-Estimation for Speaker Adaptation
Jian Wang, Jun Hai Guo, Gang Liu, Jianjun Lei · 2006
Based on speaker selection, speaker adaptation technology can get a promising performance. In such system, how to represent a speaker and the computation of selection are still big issues. In this paper, we take Gaussian mixture model (GMM) as representation of a speaker, which adapted from universal background model (UBM). Likelihood ratio (LR) and cross likelihood ratio (CLR) are utilized for speaker selection. Furthermore, a single-pass re-estimation procedure, conditioned on the speaker-independent model is shown. This adaptation strategy was evaluated in a large vocabulary speech recognition task. A relative gain of 11% with respect to the baseline system is achieved