Automatic Perceptual Evaluation of Voice Quality According to the GRBAS using Artificial Neural Networks

Abstract

In this work a comparison between two approaches for automatic classification of the GRBAS perceptual protocol of voice signals is performed. A first approach uses a classical parameterization of voice based on noise parameters and Mel frequency cepstral coefficients. In the second approach a set of parameters extracted from a nonlinear analysis of time series is used. Artificial Neural Networks have been chosen to make the classification due to the ability that they have for multi-class problems. The results show values in the fair agreement level of the Kappa index.