Application of new acoustic parameters in ANN-aided pathological speech diagnosis

Abstract

Most diseases of the vocal tract cause changes in the voice quality. Acoustic analysis of the speech signal is a widely used, noninvasive, objective and low-cost method of laryngeal pathology recognition and classification. There have been numerous attempts [1-3] to develop an automatic system which could aid the laryngological diagnosis. The goal of the presented research is to verify, whether the new approach to the acoustic analysis and parameters introduced in the Voice Analysis and Screening System (VASS 3.0 [4]) such as turbulence noise index (TNI) and normalized first harmonic energy (NFHE), can improve the effectiveness of automated diagnosis. The automated diagnosis was performed using Artificial Neural Networks (ANN). Multilayer perceptron and radial basis function neural networks of various architectures were trained to classify between pathologic and non-pathologic voices, while the parameters computed with VASS were used as input data. Preliminary results show that the Voice Analysis and Screening System coupled with ANN can be a highly effective tool for ANN-aided pathological speech diagnosis.