Modelagem de sinais de voz via PPM, aplicada ao reconhecimento de padrões vocais patológicos.

Abstract

Voice is the most widely used means of communication of mankind. However, speech organs are susceptible to several sort of pathologies, which may harm voice production and, therefore, communication. Several techniques have been used to detect these pathologies. However, they present drawbacks related to accuracy and comfort of patients during the application, which may discourage search for treatment. Thence, computational techniques have been used in order to detect the presence and type of speech pathology comfortably and accurately. But, results are still not good enough for its application in clinics, due to the fact it is considered a small number of distinct pathologies. Aiming to solve this problem, this research proposes using a method not previously employed in classification of vocal tract diseases: Prediction by Partial Matching (PPM), originally conceived for data compression purposes. The PPM model is fed with acoustical, temporal, and statistical features, ali of them extracted from voice signals. This method allowed a satisfactory classification, concerning presence and type of pathology while requiring a low computational cost (speed and storage resources). It were obtained satisfactory results regarding presence of speech pathologies. With regard to pathologies discrimination, the results suggest that this is a highly promising technique, although its application still needs deeper investigations.