Análise dinâmica não linear de sinais de voz para detecção de patologias laríngeas.

Abstract

Laryngeal pathologies may affect the voice quality, harniing human communication. The traditional objective techniques for diagnosing these pathologies make use of exams, considered invasive, causing discomfort to the patient. Acoustic analysis, using digital speech signal processing techniques. can be used for the development of non-invasive tools in order to aid laryngeal diseases diagnosis. The accuracy of diagnosis, however. depends on the choice of parameters and the speech characteristics diat better represent the voice disorder caused by a given pathology. This work deals with the characterization and classification of healthy voice signals and voices affecied by different laryngeal diseases (edema, paralysis and vocal fold nodules), by means of nonlinear dynamic analysis (and chãos theory) as well as recurrence quantification analysis. In the characterization process, the potential of each feature is investigated to discriminate the types of voice signals considered, by means of statistical tests. For the classification, the technique of discriminam analysis is employed with linear or quadratic functions, with cross-validation. A 95% confidence levei was considered for the average of accuracy rates of the classifier performance. From the feature combination of the set of nonlinear analysis measures (MNL) and the quantification recurrence measures (MQR). the average of accuracy rates varied in the following confidence intervals: [95.44%; 100%] for healthy and pathologícal classification: [94.75%; 100%] between healdiy and edema voices, and also between healthy and nodules. The accuracy rate was 100% between healthy and paralysis. We also evaluated the effects of using hybrid vectors formed by MNL, MQR and linear predictive coding (LPC) coefficients. In this case, the accuracy rates ranged in the confidence intervals: [95.02%; 97.62%] in the paralysis versus edema voices discrimination; [98.29%; 99.93%] for paralysis versus nodules and [97.98%; 99.84%] for edema versus nodules. Obtained results indicate that the used method is promising and it can even be used to develop a computational tool to support diagnosis of laryngeal diseases.