Gaussian mixture models design and applications

Two new design algorithms for estimating the parameters of Gaussian Mixture Models (GMh-l) are developed. These algorithms are based on fitting a GMM on the histogram of the data. The first method uses Least Squares Error (LSE) estimation with Gaus,s-Newton optimization technique to provide more accurate GMM parameter estimates than the commonl}' used ExpectationMaximization (EM) algorithm based estimates. The second method employs the matching pursuit algorithm which is based on finding the Gaussian functions that best match the individual components of a GMM from an overcomplete set. This algorithm provides a fast method for obtaining GMM parameter estimates. The proposed methods can be used to model the distribution of a large set of arbitrary random variables. Application of GMMs in human skin color density modeling and speaker recognition is considered. For speaker recognition, a new set of speech fiiature jmrameters is developed. The suggested set is more appropriate for speaker recognition applications than the widely used Mel-scale based one.