System of Detection and Classification of Vocal Pathologies Based on the Correntropy Spectral Density
Vocal pathologies classification, Correntropy, Multilayer Perceptron, Correntropy Spectral Density.
Vocal pathologies negatively affect the social and professional life of the sick, some of which may even lead to death if they are not treated quickly. Among the main procedures for the diagnosis of vocal disorders we can cite: laryngoscopy, perceptual-auditory evaluation, acoustic analysis of the voice, aerodynamic evaluation and self-evaluation of the voice by the patient. However, these procedures are usually invasive or inaccurate. Therefore, digital signal processing techniques have been used in the design of non-invasive systems to aid in the diagnosis of vocal tract pathologies. In this work a system of detection and classification of vocal pathologies is presented, using a classification technique based on descriptors obtained through the Correntropy Spectral Density (CSD) function, defined as the Fourier transform of the autocorrelation function. The descriptors obtained have information of second-order and higher-order statistical moments of the speech signal, through which vocal pathologies can be efficiently detected and classified. The classification is made by a neural network multilayer perceptron (MLP), performing a binary classification between normal and pathological voices and then between pathologies (edema and nodule). The classifier was evaluated by computer simulation, and the results indicate a high hit rate of detection and classification among pathologies.