Multi-Head Attention-Based Long Short-Term Memory for Depression Detection From Speech

Depression is a mental disorder that threatens the health and normal life of people. Hence, it essential to provide an effective way detect depression. However, research on depression detection mainly focuses utilizing different parallel features from audio, video, text for performance enhancement regardless making full usage inherent information speech. To focus more emotionally salient regions speech, in this research, we propose multi-head time-dimension attention-based long short-term memory (LSTM) model. We first extract frame-level store original temporal relationship speech sequence then analyze their difference between speeches those status. Then, study various use modified feature set as input LSTM layer. Instead using output traditional LSTM, attention employed obtain key time related by projecting into subspaces. The experimental results show proposed model leads improvements 2.3 10.3% over Distress Analysis Interview Corpus-Wizard Oz (DAIC-WOZ) Multi-modal Open Dataset Mental-disorder (MODMA) corpus, respectively.