Choice for a support vector machine kernel function for recognizing asphyxia from infant cries

Abstract

This paper investigates the performance of several kernel functions of support vector machine in detecting asphyxia from infant cries. In this study, Mel frequency cepstrum coefficients derived from the recorded infant cries were used as the input vectors. These input vectors were trained and classified using support vector machine. Four types of kernels - linear, quadratic, polynomial and radial basic function, were experimented and compared. Accuracy, sensitivity and specificity were adopted as criteria to obtain the best kernel. Experimental results showed that radial basic function kernel (σ = 35) is the best kernel with an accuracy of 85.15%, sensitivity of 91% and specificity of 71%.