On classification between normal and pathological voices using the MEEI-kayPENTAX database: issues and consequences

A large amount of research in pathological voice classification consider the task of feature extraction for discrimination between normal and dysphonic sustained vowels. The most widely used dataset for this purpose is the Massachusetts Eye & Ear Infirmary (MEEI) Voice Disorders Database commercialized by KayPENTAX Corp. During the last two decades, dozens of methods have been proposed to extract discriminative features from these signals in order to design accurate classifiers between the two classes of this database. The main contribution of this paper is to show that the normal and dysphonic sustained vowels of the KayPENTAX database are actually perfectly separable. This implies that this dataset is not suited for the normalvs-dysphonic classification task, as long as the only concern is to achieve high classification accuracy. Indeed, we show that a single scalar parameter extracted from a matching pursuit decomposition of these signals (with a Gabor dictionary) yields a prefect classification accuracy (100 % with a large margin). We then discuss the implication of this finding on the precaution that should be taken with this database and on research in pathological voice detection in general.