Combining multiple classifiers for age classification

Abstract

The authors compare several different classifier combination methods on a single task, namely speaker age classification. This task is well suited to combination strategies, since significantly different feature classes are employed. Support vector machines (SVMs) are trained on two different types of feature classes to estimate posterior class probabilities. The posteriors from these classifiers are combined using different combination rules and functions described in the literature. A novel age classifier is also developed by using an SVM to predict posterior class probabilities using two different types of classifier outputs; gender classification results and regression age estimates. The authors show that for combining posterior probabilities, simple combination rules such as the product rule perform surprisingly well as opposed to trainable combination strategies that require a significant amount of data and training effort