Automatic detection of voice disorders

Abstract

Speech and communication are the bases of our society and the quality of the speech can seriously affect any person’s life. Besides the irregularities in the voice production can be caused by different diseases which can be treated better if they are diagnosed in an early stage. In this research we introduce a series of measurements based on continuous speech which could be a good base of developing a system that can automatically distinguish several voice disorders and identify the quality of the voice according to widely used RBH scale ((Rauhigkeit) (roughness) (Behauchtkeit) (breathiness) (Heiserkeit) (hoarseness)): 0 = normal voice quality, 3 = heavy huskiness). The different feature extraction and classification experiments presented in this research show how it is possible to separate healthy from pathological speech automatically and demonstrate possibilities to define the type of the pathological voice. The results suggest that the automatic classification into two or multiple classes can be improved by a multi-step pre-processing methodology, in which only the most significant acoustic features of a given class are extracted from the voice and used to train the SVM (Support Vector Machine) classifiers. We also present the importance of data acquisition, and how the selected features and the number of training samples can affect the accuracy and performance of the automatic voice disorder detection.