Pathological Voice Classification Based on Wavelet Packet Multiscale Analysis

Vocal fold paralysis is a common type of laryngeal diseases, and it is fundamental different from vocal fold non-paralysis (including vocal fold nodules, polyps, etc.). Based on the differences between the various of laryngeal diseases, a classification algorithm for pathological voice is proposed. In which the wavelet packet transformation and multi-scale analysis are used. Firstly, the sub-bands of different frequency signals are obtained using wavelet packet decomposition, and then the nonlinear features i.e. the Hurst parameter, 2-Renyi entropy, Box-counting dimension and attractor are extracted from different frequency bands. These features are used to evaluate the contribution of each frequency band in detecting and classifying pathological voices. The experimental data are derived from the Massachusetts Eye and Ear Infirmary (MEEI) database and the Saarbrucken Voice Database (SVD). The experimental results show that when using the support vector machine to classify the Hurst parameter and the 2-Renyi entropy combined features, the method used in this paper can achieve good classification results when classifying normal, paralysis and non-paralysis these three kinds of voices on those two databases. The average classification accuracies on these two databases are 98.37% and 92.83% respectively.