Skip to Main Content
The session variability is the most important factor affecting the performance of the speaker verification systems. In order to deal with the variability more efficiently, this paper provides a practical procedure for applying a smooth within-class covariance normalization (WCCN) to an SVM-based speaker verification system, where the dimension of input samples resides in a low session-invariant principal component analysis(SIPCA) feature space. When the SIPCA and smooth WCCN approaches are implemented on NIST 2006 verification task, experimental results show relative reductions of up to 19.7% in EER and 18.4% in minimum decision cost function(DCF) over our previous GMM-mean SVM system. Our approach also has advantages in computational and memory costs compared to the state-of-art systems.