Combining Gaussian Mixture Models and Polynomial Classifiers for Text Independent Speaker Recognition

摘要

A framework for combining polynomial classifiers with Gaussian mixture models is given. This framework is based upon earlier work combining hidden Markov models and polynomial classifiers for text-prompted speaker recognition. The basic idea is to use a Gaussian mixture model as a decoder and then perform final scoring with a polynomial classifier. Results for both support vector machine training and minimum mean-squared error training are given. Experiments on the NIST 1998 speaker recognition database show the viability of the method.