Speaker Recognition using Excitation Source Parameters

Abstract

Excitation signal is used in speaker recognition. It corresponds to the frequency of oscillation of vocal cords and is one of the speaker's characteristics. Although this feature gives worse recognition results compared to the vocal tract parameters, but it is more robust to various distortions in the recording channels. As a result, pitch is commonly used in forensic investigations, where different recording channels is one of the main problems. Currently, the pitch distribution generally is modeled using histograms and calculating various distances or similarity measures between two histograms. However, pitch distribution is not Gaussian and view of the histograms and comparison results depend on the number of classes used. We model pitch distribution using Gaussian mixture models (GMM), and calculate similarity and distance measures between the GMM approximations of two comparative records. Best results were achieved using symmetric Kullback-Leibler distance.