TDCA-Net: Time-Domain Channel Attention Network for Depression Detection

Depression is a psychiatric disorder and has many adverse effects on our society. Some studies have shown that speech signals are closely related to emotion and stress, and many speech-based automatic depression detection methods have been proposed. However, previous work is based on spectrogram or hand-crafted features, which may lose some useful information related to depression patterns. And there is no evidence that the filter bank designed from perceptual evidence is optimal for depression detection. In order to learn the more discriminative feature representation related to depression, we propose an end-to-end time-domain channel attention network (TDCA-Net) for depression detection. The TDCA-Net directly models time-domain speech signals based on dilated convolution block, which can increase the receptive field exponentially and aggregate multiscale contextual information associated with depression. Besides, we employ the efficient channel attention (ECA) module to model dependencies of channels and improve the sensitivity of the model to information related to depression. Experimental results on the AVEC2013 and the AVEC2014 datasets illustrate the effectiveness of our method.