Simplified factor analysis in speaker verification

The channel or session variability is one of the major difficulty problems to deal with in speaker recognition systems. In this paper, a simplified factor analysis is proposed to solve the inter-session variability problem. We find the factor analysis method can achieve very good performance both in the GMM_UBM and the SVM systems. The proposed algorithm can get almost the same performance as the NAP algorithm in SVM system. In the NIST 2006 SRE core test, the EER of the factor analysis is about 5.0% in the SVM system and 4.9% in the GMM_UBM system.