Application of differential evolution optimization based Gaussian Mixture Models to speaker recognition

Voice-based speaker recognition technique can be used in the identification of speakers. In such manner, Gaussian Mixture Model (GMM) can provide voice feature vectors´ probability density model. In this paper, the Akaike´s Information Criterion (AIC) is used to identify structures of the GMM models. The GMM parameter optimization is done by the differential evolution (DE) algorithm. During the optimization, a new parametric method is applied aiming at ensuring the positive definite symmetry property of an arbitrary covariance matrix. Here, both the expectation-maximization (EM) and DE are applied to identify the GMM parameters of a simulated dataset, and the utility of DE is proved by comparing the performances of the two. Further, DE is used to identify parameters of the GMM of Speaker Dataset acquired by Information Processing Laboratory in Hokkaido University. Again, the good performances of DE demonstrate superiorities to the EM method.