Feature selection using adaptive learning networks for text-independent speaker verification

Abstract

A nonlinear transformation which makes use of the polynomial discriminant function (PDF) is applied to the selection of features for text-independent speaker verification. For each speaker within the set, a PDF is established such that it maps the original attribute set onto the one-dimensional one resulting in the best discrimination of the speaker from the others in the new feature space. The complicated PDF is implemented using the technique known as the adaptive learning network (ALN). In this case, the multinomial discriminant function is estimated using interconnecting fundamental building blocks each of which computes a two-element quadratic discriminant function. Through a comprehensive training procedure, the coefficients needed to describe each building block and the interconnecting configurations of them are determined. Application of the chosen attributes to a text-independent speaker verification experiment yields relatively low false speaker rejection and verification error rates.