Evaluation of SVM Kernels and Conventional Machine Learning Algorithms for Speaker Identification

One of the central problems in the study of Support vector machine (SVM) is kernel selection, that’s based essentially on the problem of choosing a kernel function for a particular task and dataset. By contradiction to other machine learning algorithms, SVM focuses on maximizing the generalisation ability, which depends on the empirical risk and the complexity of the machine. In the following paper, we considered the problem of kernel selection of SVMs classifiers to achieve performance on text-independent speaker identification using the TIMIT corpus. We were focused on SVM trained using linear, polynomial and Radial Basic Function (RBF) kernels. A preliminary study has been made between SVM using the best choice of kernel and three other popular learning algorithms, namely Naive Bayes (NB), decision tree C4.5 and Multi Layer Perceptron (MLP). Results had revealed that SVM trained using polynomial kernel is the best choice for dealing with speaker identification tasks and that SVM is the best choice when compared to other algorithms.