An anticorrelation kernel for improved system combination in speaker verification

This paper presents a method for training SVM-based classification systems for combination with other existing classification systems designed for the same task. Ideally, a new system should be designed such that, when combined with the existing systems, the resulting performance is optimized. To achieve this goal, we include a regularization term in the SVM objective function that aims to reduce the within-class correlation between the resulting scores and the scores produced by one of the existing systems, introducing a trade-off between such correlation and the system’s individual performance. That is, the SVM system “takes one for the team”, falling somewhat short of its best possible performance in order to be more complementary to the existing system. We report results on the NIST 2005 and 2006 speaker recognition evaluations (SRE) using three component systems: a standard UBM-GMM system, an MLLR-based system, and a prosodic system, and show that the proposed technique results in performance gains of 16% in EER and 23% in DCF.