Speaker Verification Experiments on the XM2VTS Database

This paper describes two speaker verification algorithms, a text-independent method based on a second order statistical measure and a text-dependent method based on hidden Markov modelling. We investigate the effect of different features, sampling rates, and threshold setting methods and introduce a N-best words pruning method that aims to compensate for the effect of poorly trained client models. Experimental evaluation is performed on the publicly available XM2VTS database according to a published protocol for three different operating points and a priori threshold setting