Pathological Voice Classification Based on Recurrence Quantification Measures

Abstract

This paper presents an analysis of speech signals based on quantification measures of recurrence plots. A comparison between healthy voices and voices affected by laryngeal pathologies (Reinke's edema, nodule and vocal fold paralysis) is made. In order to classify these signals as pathological or healthy, seven recurrence quantification measures are used: Determinism, maximum length of the diagonal structures, Shannon entropy of line distribution, slope of line of best fit, laminarity, length of longest vertical line segment and mean vertical line length or trapping time. Discriminant analysis methods (linear and quadratic) are applied to each feature individually, and to the vectors formed by feature combination with cross-validation classification rates up to 95.71±4.94% (95% confidence interval). Results show that the employed measures present a significant discriminative potential to distinguish healthy voices from pathological ones.