RC-MES: a novel speaker modeling technique based on regression class for speaker identification

摘要

The speaker modeling technique is an essential problem in robust speaker recognition, especially when enrolment data is sparse. This paper presents a novel modeling approach named multi-eigenspace modeling technique based on regression class (RC-MES), which integrates the common eigenspace technique and the regression class (RC) idea of maximum likelihood linear regression (MLLR). RC-MES not only solves the problem of prior knowledge limitation of Gaussian mixture models (GMM) but also remedies the shortcomings of the common eigenspace that confuses speaker differences and phoneme differences. The eigenvoice analysis in RC can provide better discrimination ability between different speakers. The experimental results on speaker identification of 75 males show that, when enrolment data is sparse, RC-MES provides significant improvement over GMM, and the number of eigenvoices in RC-MES is fewer than that in the common eigenspace.