Data Augmentation in Automatic Classification of Voice Quality

Abstract

The status of the thesis plan "Vocal quality assessment through deep scattering spectrum and machine learning" is presented. Three transformations are proposed in order to increase the amount of training data and reduce overfitting. These transformations perform a frequency shift, time segmentation and flipping. It results in a dataset 18 times larger than the original dataset. An experiment consisting of training a deep neural network is run to evaluate performance with the augmented data. It is concluded that the proposed transformations reduce the overfitting, improve the classification error and it could be useful for the thesis plan scope, classification of vocal quality from sustained vowels.