Regression optimized kernel for high-level speaker verification

Computing the likelihood-ratio (LR) score of a test utterance is an important step in speaker verification. It has recently been shown that for discrete speaker models, the LR scores can be expressed as dot products between supervectors formed by the test utterance,"target-speaker model, and background model. This paper leverages this dot-product formulation and the representer theorem to derive a general kernel, namely the regression optimized kernel, for computingutterance-based verification scores using support vector machines."The kernel is general in that it can be a linear combinationof any kernels belonging to the reproduction kernel Hilbert space. The combination weights are obtained by maximizing the ability of a discriminant function in separating a target speaker from impostors. The regression optimized kernel was applied to high-level speaker verification using articulatory-feature based pronunciationmodels. Results show that the scores produced by the regression optimized kernel are not only superior but also complementary to the LR scores, resulting in better performance when the two types of"scores are combined. The proposed regression optimized kernel canbe easily applied to other SVM-based classification problems.