Analysis of Speech Based on Spectral Entropy in Detecting Depressed among Control Subjects

This paper presents the study on entropy in speech for its possible vocal predictor of severe depression in speakers. Persons with recurring depressive symptom are possibly at risk of suicide when the symptom strikes, unless admitted into hospital and have proper treatment in time. Prediction is primarily necessary task to prevent such risk. In this study the full-band and further sub-band entropies of eight evenly separated frequency bands of 625 Hz estimated from the female voiced segments were extracted and consequently used to form the parameter models for the betweengroup classifications. The average of correct classification is fairly high when training ML classifier with 20% of extracted entropy samples and then evaluating it later with the rest of same database and comparing to cases of training ML with different sizes of separated samples in trainings. As result shown, the obtained classifying percentages suggest the sub-band entropy in speech signal capable of separating between two categorized speaker groups.