Comparison of current frame-based phoneme classifiers

Abstract

This paper compares today’s most common frame-based classifiers. These classifiers can be divided into the two main groups – generic classifiers which creates the most probable model based on the training data (for example GMM) and discriminative classifiers which focues on creating decision hyperplane. A lot of research has been done with the GMM classifiers and therefore this paper will be mainly focused on the frame-based classifiers. Two discriminative classifiers will be presented. These classifiers implements a hieararchical tree root structure over the input phoneme group which shown to be an effective. Based on these classifiers, two efficient training algorithms will be presented. We demonstrate advantages of our training algorithms by evaluating all classifiers over the TIMIT speech corpus.