Clustering based under-sampling for improving speaker verification decisions using AdaBoost

Abstract

The class imbalance problem naturally occurs in some classification problems where the amount of training samples available for one class may be much less than that of another. In order to deal with this problem, random sampling based methods are generally used. This paper proposes a clustering based sampling technique to select a subset from the majority class involving much larger amount of training data. The proposed approach is verified in designing a post-classifier using AdaBoost to improve the speaker verification decisions. Experiments conducted on NIST99 speaker verification corpus have shown that in general, the proposed sampling technique provides better equal error rates (EER) than random sampling.