Simplified channel factor estimates in speaker verification
Ren-Hua Wang · Journal of Tsinghua University(Science and Technology) · 2008
Factor analysis is one of the most effective methods to reduce the channel bias in speaker-independent speaker verification methods.The factor analysis training requires a large speech corpus and a complex EM(expectation maximization).Furthermore,the log-likelihood ratio calculation used in the factor analysis test process prevents real-time applications.In this work,the traditional principal component analysis(PCA) method was used with feature mapping to reduce the factor analysis work load.The channel subspace is first estimated using GMM(Gaussian mixture model) supervectors,then the channel bias is subtracted from the speech acoustic features.With this method,the equal error rate(EER) in the NIST 2006 SRE 1conv4w-1conv4w corpus is reduced by 24% compared to the baseline GMM system.The algorithm reduces the computations in the speaker verification system while maintaining a high recognition rate.