Nonlinear decision function in speaker verification using a classifier ensemble

The decision rule in speaker verification systems depends on a linear Bayes decision boundary which can be controlled with a threshold. In this paper, the use of complex and nonlinear boundary based decision making is explored which can be achieved using multiple classifier approach. The potential problems in applying such techniques in speaker verification are specified together with some candidate solutions. Then, a well known boosting technique called AdaBoost which is effective in creating an ensemble of classifiers is described. Experiments conducted on NIST99 speaker verification corpus has shown that nonlinear boundary obtained using AdaBoost provides 9.2% improvement in the equal error rate (EER) compared to the Bayes decision making.