Multi-Method System for Robust Text-Independent Speaker Verification

A system for robust speaker verification based on four recognition approaches and methods (classifiers) is proposed, in order to use different statistical characteristics of the speech parameters. These methods are: 1) Prototype Distribution Maps (PDM); 2) AR-vector models (ARVM); 3) Two-level approach: the first level uses several PDMs for preprocessing, and the second employs multilayer perceptron (MLP) networks; 4) Gaussian speaker’s models combined with the arithmetic-harmonic sphericity measure (GMAHSM). These classifiers generate four preliminary classification decisions. The reliability and confidence of these preliminary decisions are evaluated by means of a weighting algorithm. The weights are assigned using the relative measures to the most similar speakers (or cohorts), i.e. a Cohort Normalization technique is implemented. The final classification is then performed using simple logical and threshold rules. The speech signals of 92 speakers have been analyzed. The speaker verification accuracy was over 98%. Robust impostors’ detection was observed, because the classifiers never fail simultaneously.