Articulatory-feature based sequence kernel for high-level speaker verification

Research has shown that articulatory feature-based phonetic-class pronunciation models (AFCPMs) can capture the pronunciation characteristics of speakers. However, the scoring method used in AFCPMs does not explicitly use the discriminative information available in the training data. To harness this information, this paper proposes converting speaker models to supervectors by stacking the discrete densities in AFCPMs. An AF-kernel is constructed from the supervectors of target speakers, background speakers, and claimants. An AF-kernel based SVM is then trained to classify the super-vectors. Results show that AF-kernel scoring is complementary to likelihood-ratio scoring, leading to better performance when the two scoring methods are combined.