Automatic Voice Quality Measurement Based on Efficient Combination of Multiple Features

Abstract

This work proposes higher-order statistics (HOS)- based features to improve classification performance of voice quality measurement. They are means and variances of skewness and kurtosis which show meaningful differences in normal, breathy, and rough voices. Jitter, shimmer, and harmonic to noise ratio (HNR) are implemented as conventional features. The performances are measured by classification and regression tree (CART) analysis. Specifically, the CART-based method by utilizing both conventional and HOS-based features is shown to be an effective for voice quality measurement, with an 89.7% classification rate.