On the optimization of SVM kernel parameters for improving audio classification accuracy

This paper presents a grid search approach to optimize the kernel's parameters for the support vector machines classifier. The most encountered three kernels are considered: linear, radial basis, and sigmoid. We show that the optimization of parameters improves the recognition performance for audio signals classification, especially in the case of sigmoid kernel. The behavior of the model is very sensitive to kernel's parameters, which, in turn are sensitive to data and selected features. We consider the problem of multiclass classification with imbalanced datasets. We compare the accuracies obtain with and without kernel parameter's optimization. As features we use Mel frequency cepstral coefficients.