Supervised neural network classifier for voice pathology

Abstract

The classification of normal and pathological voices using noninvasive acoustical analysis features helps speech specialist to perform accurate diagnoses of vocal and voice disease. Acoustic analysis is a non-invasive technique based on digital processing of the speech signal. Ear Nose and throat (ENT)clinicians and speech therapists uses subjective techniques or invasive methods such as evaluation of voice quality by the specialist doctor's direct inspection and the observation of vocal folds by endoscopy techniques for the evaluation and diagnosis of voice pathologies. These Techniques provide inconvenience to the patient and depend on expertise of medical doctors. In the evaluation of quality speech, acoustic analyses of normal and pathological voices have become increasingly interesting to researcher in laryngology and speech pathologies. This paper present a new measure to parameterize the voice signal based on energy levels extracted from each frame of speech signal. A supervised neural net classifier for the classification of pathological voices using Back propagation with variable slope parameter is proposed. a simple scheme is proposed to fix the slope parameterof the bipolar/binar sigmoidal activation function. Simulation results indicate that the proposed classification algorithm distinguish the voice as pathological or a non-pathological voice accurately.