An Enhanced Centered Binary Tree of SVMs Algorithm for Phoneme Recognition

Support vector machines (SVMs) have lately emerged as very powerful binary classifiers, however their extension for multiclass classification is still an open area of research. When SVMs are applied to phoneme recognition, the large number of classes within this data set limits the practical use of these learning machines. In this paper we present the application of the centered binary tree of SVMs (c-BTS) algorithm, which is a multi-category classifier, to phoneme recognition. To enhance its capabilities, the c-BTS algorithm has been modified by using different posterior probability measurements to build the binary tree. The proposed algorithm has been tested through simulations on a number of phonemes taken from the TIMIT database. Results show that the proposed modification enhances the accuracy of the c-BTS algorithm while maintaining comparable computation times during testing. Moreover, this solution offers equivalent accuracies to other multiclass recognition methods.