The Turkish Audio-Visual Bipolar Disorder Corpus

Abstract

This paper introduces a new audio-visual Bipolar Disorder (BD) corpus for the affective computing and psychiatric communities. The corpus is annotated for BD state, as well as Young Mania Rating Scale (YMRS) by psychiatrists. The paper also presents an audio-visual pipeline for BD state classification. The investigated features include functionals of appearance descriptors extracted from fine-tuned Deep Convolutional Neural Networks (DCNN), geometric features obtained using tracked facial landmarks, as well as acoustic features extracted via openSMILE tool. Furthermore, acoustics based emotion models are trained on a Turkish emotional database and emotion predictions are cast on the utterances of the BD corpus. The affective scores/predictions are investigated with linear regression and correlation analyses against YMRS declines to give insights about BD, which is directly linked with emotional lability, i.e., quick changes in affect.