Emotional and behavioral disorders (EBD) are a widespread healthcare concern in children and adolescents. Prevention and early intervention are the most powerful tools in ameliorating the problem, and therefore, timely and accurate detection of abnormal emotional patterns is of vital importance. In this paper, we propose a system that detects second-level emotional states of children using hour-level audio recordings. The proposed system consists of an audio segmentation and speaker tracking front-end along with an emotion recognition back-end. Supervised support vector machine is used in the front-end to improve its robustness to short and inconsistent child speech pattern and end-to-end deep learning is used in the emotion recognition back-end to improve its robustness to noise and segmentation error. We further demonstrate the potential of the proposed system as an automated emotion analysis tool.
