Inteligência artificial aplicada no auxílio à detecção de patologias vocais.

Abstract

Voice listening is a way of assessing vocal health, in which a professional judges the patient's voice as pathological or not after listening to it. The problem with this method is its subjective nature, since the result can vary depending on the examiner. For a more accurate analysis, laboratory techniques can be applied; however, they are often avoided by patients due to their invasive and costly nature. Thus, researchers have developed techniques to assist in the discrimination of pathological voices using acoustic analysis, as it is a non-invasive and automatic form of digital signal processing. This method consists of using digital signal processing and pattern recognition techniques to determine whether the voice signal is pathological or not. In view of this, this article aims to analyze the use of an artificial neural network (ANN) as a classifier and characteristics obtained through Mel Cepstral Coefficients (MFCC), which will assist in the detection of voice pathologies. For the training and validation of the ANN, the German database Saarbruecken Voice Database (SVD) was used. The results demonstrated, with k-fold cross-validation for training and testing, that the solution achieved accuracy levels above 85% in distinguishing between healthy and pathological voices.