Classificação de sinais vocais em parâmetros não acústicos utilizando redes neurais artificiais

Abstract

This work proposes an aid method for speech professionals, in clinical, scientific and pedagogical environments, since the classification of vocal samples into subjective parameters is extremely important for the diagnosis of pathologies and much used in the daily life of these professionals. Using previously recorded samples, the patient's voice is analyzed by means of acoustic parameters - energy and logarithmic entropy - extracted from the third level coefficients of the Wavelet Packet Transform. Subsequently these samples are classified by a set of artificial neural networks in roughness, breathiness or strain, all subjective parameters of the voice.In order to obtain a higher rate of accuracy, the study proposes the use of six artificial neural networks, each one specialized in the identification of one of the subjective parameters, also divided into female and male.The experimental results demonstrate that the proposed methodology can adequately generalize speech samples with a mean percentage of correctness of 96.33%.