An Alternative To Silence Removal For Text-Independent Speaker Verification

State-of-the-art text independent speaker verification systems use silence/speech detectors to get rid of silence frames which are considered to be non discriminative. This paper explores a possible replacement to this silence/speech detector by considering each Gaussian of a GMM as modeling a specific speech class and by using discriminant models like SVMs and MLPs in order to fuse the corresponding class-specific scores to obtain a final decision. Experiments on the NIST 2001 database yielded statistically significantly better performance for the new model as compared to our best baseline system involving a silence/speech detector, without having to rely on uncertain hypotheses.