Discrimination between pathological voice categories using matching pursuit

Abstract

There are several methods in the literature for pathological voice classification but there are very few methods which can classify pathological sub-groups. An attempt is made here to classify pathological sub-groups using matching pursuit decomposition method and is compared with PRAAT. Random forest classifier is used and frequency band of the atoms are used as feature. The result shows that we can classify adductor spasmodic dysphonia, keratosis and vocal nodules in a class of voices consisting of adductor spasmodic dysphonia, keratosis, paralysis, vocal nodules and vocal fold polyps with reasonably good classification accuracy. Both matching pursuit (MP) and PRAAT shows comparable classification scores but using MP is more advantageous over PRAAT since it doesn't rely on pitch information and extraction of pitch information in a pathological signal is a complex problem.