A time-frequency channel attention and vectorization network for automatic depression level prediction

Physiological studies have illustrated that speech can be used as a biomarker to analyze the severity of depression and different frequency bands of the speech spectrum contribute unequally for depression detection. To this end, we propose a Time-Frequency Attention (TFA) component and combine it with the Squeeze-and-Excitation (SE) component to form our Time-Frequency Channel Attention (TFCA) block for emphasizing those discriminative timestamps, frequency bands and channels. In addition, considering the time-frequency attributes of the data, a Time-Frequency Channel Vectorization (TFCV) block is proposed to vectorize the tensor. Furthermore, we merge the proposed blocks (i.e., TFCA and TFCV blocks) and the two blocks (i.e., Dense block and Transition Layer) of the DenseNet into a unified architecture to form our Time-Frequency Channel Attention and Vectorization (TFCAV) network. In this way, to predict the depression level of an individual, we firstly introduce the sphere embedding normalization method to preprocess the long-term logarithmic amplitude spectrum for maintaining the time frequency attributes and divide it into segments. Then, these segments are input into the TFCAV network to obtain the depression scores. Finally, the average of scores is taken as the result corresponding to the long-term spectrum. Our method is validated on two challenging databases, i.e., AVEC2013 and AVEC2014 depression databases. The experimental performance illustrates the superiority of the proposed network over some previous methods. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

This study proposes a new method for depression detection by combining time-frequency attention and squeeze-and-excitation components, emphasizing discriminative timestamps, frequency bands, and channels. Through attention and vectorization processing of the time-frequency attributes of data, a new network architecture is formed, improving the accuracy and efficiency of depression detection.