A comparative study of statistical ensemble methods on mismatch conditions

Abstract

Unlike previous comparative studies, we present an empirical evaluation on three typical statistical ensemble methods - boosting, bagging and combination of weak perceptrons - in terms of speaker identification where miscellaneous mismatch conditions are involved. During creating an ensemble, moreover, different combination strategies are also investigated. As a result, our studies present their generalization capabilities on mismatch conditions, which provides an alternative insight to understand those methods.