Towards a small intra-speaker variability models

Abstract

Automatic speaker recognizer used in experiments described in this paper uses vectors of melfrequency cepstral coefficients as feature vectors and covariance matrices for speakers modelling. By comparing the models of training and test speech of the same speakers it was noticed significant differences in some model elements. Speaker models had inherent intra-speaker variability. In the observed test the distinction matrix was introduced as the measure of intra-speaker variability of all speaker models in available speech database. Based on the values of elements in distinction matrix, the ranges of validity of elements in weighting matrix were established. Each element in covariance matrix of a speaker was pondered by appropriate weighting coefficient. Application of this transformation resulted in higher accuracy of automatic speaker recognition.