Depression Detection by Person`s Voice

Abstract

In this work, a machine learning algorithm is proposed to detect depression. The Transformer encoder network is considered and compared with top baseline approaches. Low-level features are extracted from audio recordings and then are augmented to overcome the problem of the small size of available dataset. The Transformer network achieves recognition accuracy of 73.51% on DAIC-WOZ database, which compare favourably to the accuracy of 65.85% and 66.35% obtained by traditional approaches.