Classificação de sinais de voz através da aplicação da transformada Wavelet Packet e redes neurais artificiais

Abstract

Pathologies such as edema, nodules and paralysis are quite recurrent and directly influence vocal dysfunctions. The acoustic analysis has been used to evaluate the disorders caused in the voice signals, detecting the presence of pathologies in the larynx, through digital signal processing techniques. This work aims to distinguish healthy voice signals from the ones affected by laryngeal pathologies, using the Wavelet Packet transform in the feature extraction step. Energy and entropy measures, in six resolution levels, obtained through the Daubechies wavelet of order 4 are used in the discrimination of the voice signals. The classification is done through Artificial Neural Networks. Accuracies above 90% were obtained, with the entropy measure, in the discrimination between healthy voices and affected ones by pathologies in the vocal folds (nodules, Reinke’s edema and paralysis)