Subdividing of pathological voice based on vocal tract area

Abstract

Vocal tract is one of the most important system in speech production. Vocal tract shape (VTS) is defined as varying cross sectional area from glottis to lips. Normal and pathological voice will show different VTS and even different kinds of voice diseases will exhibit various VTS with the same vowel sound. In this work, we propose a method to verify the view numerically. We model the vocal tract as a multi-sectional cylindrical tube, with each section having the same length and a different cross-sectional area, then we estimate the vocal tract area (VTA) of /a/ sound from Soochow University Pathology Voice Database over the closed glottal phase using the attenuated weighted linear prediction method. Features are extracted from VTA which has 70 dimensions. The vocal tract features and three kinds of classifiers are used in this study, the best subdividing accuracy can achieve 97.5% which is better than many methods in the area of subdividing different voice diseases and indicates the view we proposed early.