Classificação de disfonias utilizando redes neurais artificiais e transformadas Wavelet Packet

Abstract

Most laryngeal pathologies cause deviations in the vibratory pattern of the vocal folds. These deviations produce changes in vocal quality and can become a significant problem, especially for occupational and professional voice users. Voice assessment methods are generally invasive, have a high implementation cost and cause discomfort to the patient. On the other hand, noninvasive methods do not have good resolution and robustness, since they assume voice as a periodic signal, without considering that the more pathological the voice, the more aperiodic is the signal. Artificial Neural Networks as Multilayer Perceptron have been widely applied in the recognition of nonlinear biological signal patterns, presenting good accuracy, with high reliability. For these reasons, this work proposes the application of Artificial Neural Networks and Wavelet Packet Transform for voice signal analysis and for classification of functional and organic dysphonias, as well as healthy voices. In this way, 100% classification accuracy was obtained for healthy, 99.55% for organic dysphonia and 93.47% for functional dysphonia, with 98% reliability.