TEXT-INDEPENDENT SPEAKER VERIFICATION USING AUTOMATICALLY LABELLED ACOUSTIC SEGMENTS

Most of the current text-independent speaker verification techniques are based on modelling the global probability distribution function of speakers in the acoustic vector space. We present an alternative approach based on class-dependent verification systems using automatically determined segmental units, obtained with temporal decomposition and labelled through unsupervised clustering. The core of the system is a set of multi-layer perceptrons (MLP) trained to discriminate between client and an independent set of world speakers. Each MLP is dedicated to work with data segments that are previously selected as belonging to a particular class. Issues and potential advantages of the segmental approach are presented. Performances of global and segmental approaches are tested on the NIST'98 database (250 female and 250 male speakers), showing promising results for the proposed new segmental approach. Comparison with a state of the art system, based on Gaussian Mixture Modelling is also included.