Malayalam language vowel classification using Support Vector Machine for children

This work intended to investigate the performance of the support vector machine (SVM) classifier in the problem area of Automatic vowel Recognition in the Malayalam monophthongs vowel corpus of children in the age group of five to ten. Pattern analysis and classification of children’s speech, based on their acoustic features, are much more complicated than that of adults. In most cases, the inadequacy of the regional language speech corpus of children is the major constraint of the research and development of this area. SVM can work effectively even in small training datasets. The mel frequency cepstral coefficient and its derivatives, Formants (F1, F2, and F3), and zero crossing rate (ZCR) are the features selected by the feature selection method for this problem. This work analyses the performance of the SVM with Quadratic (88%), Cubic (89.5%), Fine Gaussian (77.5%), Medium Gaussian (91.5%), and Coarse Gaussian (82%) kernel functions. The Bayesian optimization method and Grid search optimization method used for tuning the hyper-parameter and optimized model outperformed all the other mentioned basic non-linear SVM models. The Bayesian optimizer has shown the best performance with Cubic SVM (91.5%), and Grid search optimizer has shown the best performance with Quadratic SVM (93%).