Performance of Acoustic Measures for the Discrimination Among Healthy, Rough, Breathy, and Strained Voices Using the Feedforward Neural Network.

Abstract

Objective: To identify and evaluate the best set of acoustic measures to discriminate among healthy, rough, breathy, and strained voices.

Methods: This study used the vocal samples of the sustained /ε/ vowel from 251 patients with the vocal complaints, among which 51, 80, 63, and 57 patients exhibited healthy, rough, breathy, and strained voices, respectively. Twenty-two acoustic measures were extracted, and feature selection was applied to reduce the number of combinations of acoustic measures and obtain an optimal subset of measures according to the information gain attribute ranking algorithm. To classify signals as a function of predominant voice quality, a feedforward neural network was applied using a Levenberg-Marquardt supervised learning algorithm.

Results: The best results were obtained from 11 combinations, with each combination presenting six acoustic measures. Kappa indices ranged from 0.7527 to 0.7743, the overall hit rates are 81.67%-83.27%, and the hit rates of healthy, rough, breathy, and strained voices are 74.51%-84.31%, 78.75%-90.00%, 85.71%-98.41%, and 68.42%-82.46%, respectively.

Conclusions: We obtained the best results from 11 combinations, with each combination exhibiting six acoustic measures for discriminating among healthy, rough, breathy, and strained voices. These sets exhibited good Kappa performance and a good overall hit rate. The hit rate varied between acceptable and good for healthy voices, acceptable and excellent for rough voices, good and excellent for breathy voices, and poor and good for strained voices.