Speaker-Independent Depression Detection based on Adversarial Training Method

Depression, a worldwide mental health problem, significantly affects people's lives and may even lead to suicide in severe cases. As speech contains a large number of features that can reflect the speaker's mental condition, research has been conducted on depression detection based on speech analysis recently. However, personalized features from various speakers in speech can be mingled with depression-related features, causing a personalization bias in the features and reducing feature generalization. In this paper, we propose an adversarial training method based on the gradient reversal technique to remove speaker characteristics from speech signals, eliminating the personalization bias caused by speaker characteristics on depression features. Experiments carried out on the Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WoZ) dataset, demonstrate that our method improved the detection performance through more generalized depression feature extraction.