Multi-Head Self-Attention Network for Depression Level Estimation from Speech

Depression levels estimation from speech is a hot research topic currently. Researches have shown that speech signals are intimately related to stress and emotions. Therefore, the researchers focused on extracting features associated with depression from the speech signals. For example, Mel spectrograms or Fourier spectrograms are utilized to find clues related to depression, but the results are not satisfactory. We suggest modeling the raw speech signal directly and propose a WavMHSA-Net. Specifically, we use the WavLM base model to extract features directly from speech signals. Then we utilize the multi-head self-attention(MHSA) mechanism to highlight the features related to depression and establish the global correlation. Finally, we propose a prediction score module that vectorizes features to predict the final depression level score (BDI-II score). The effectiveness of the method in this paper is demonstrated by experiments on publicly available depression detection datasets (AVEC2013 and AVEC2014).