Detecção de nódulos nas pregas vocais através da proeminência do pico cepstral de vogal sustentada

Abstract

This work aims to detect nodules in the vocal folds through the prominence of the cepstral peak (CPP) and the prominence of the smoothed cepstral peak (CPPS) of a sustained vowel. As in the literature, the power cepstr was calculated on the decibel scale and, for comparison purposes, this work calculated the cepstr on the linear scale. 10 descriptive statistics from 8 CPP and 12 CPPS configurations for 35 volunteers were considered. To assess the discriminant capacity of descriptive statistics, the Kruskal Wallis test was performed. The CPP mode calculated on the decibel scale with windowing time T = 20ms showed the best discriminant capacity, reaching a p-value of 4.22e-06. The classification of statistics was performed using a leave- one-out cross-validated SVM and considering three different types of kernels: linear, polynomial and Gaussian. The best accuracy obtained was 0.97 for statistical combinations mean/fifth percentile/asymmetry and median/fifth percentile/asymmetry of the CPP calculated on the decibel scale with windowing time T = 20ms. Overall, the results indicate that the combination of statistics related to the location of the distribution with statistics that characterize the shape of the distribution are promising for classification. Although the literature affirms the efficiency of the CPPS to distinguish healthy from pathological voices, in this work the CPP presented a better discriminating ability in the Kruskal Wallis Test and better accuracy values in the classification.