Non-Stationarity-Based Adaptive Segmentation Applied to Voice Disorder Discrimination

Abstract

Voice disorders can affect people in all age groups. Vocal health is especially important when voice is used as a work tool, as in the fields of education (teachers), music (singers) and multimedia (journalists). Several researches based on linear and nonlinear models of speech production have attempted to identify the intrinsic characteristics in voices affected by pathologies of organic or functional origin. However, there has been no consensus thus far on the best measure to distinguish a healthy voice from a pathological voice, whether in the diagnosis or treatment phase of a disorder. A common procedure in feature extraction uses a traditional voice segmentation, based on stationary frame sizes. On the other hand, no studies have attempted to analyze how laryngeal pathologies can introduce non-stationary acoustic variations in voice. In this study, the Index of Non-Stationarity (INS) is used to identify frame sizes that can be considered truly stationary for each signal. Healthy voice signals and voice signals affected by pathologies such as paralysis, Reinke's edema and nodules are analyzed, and a comparative analysis of the traditional and INS-based adaptive segmentation is performed. Linear predictive coding (LPC) coefficients, melfrequency cepstral coefficients (MFCC), gammatone-frequency cepstral coefficients (GFCC) and recurrence quantification measures (RQMs) are used as acoustic measures. Classification is carried out by performing a linear discriminant analysis (LDA). The results show that the proposed adaptive segmentation can increase pathology recognition (sensitivity) by up to 18%, for the linear model of speech production.