MDSNet: Exploring Refined Speech Representations for Screening Mental Disorders

Speech-based mental disorder screening (SMS) is hindered by the scarcity of labeled training data. Although self-supervised learning (SSL) can alleviate this limitation by leveraging unlabeled speech, the general representations learned by existing SSL methods often contain unhelpful redundancies and are not well adapted to SMS tasks. To solve these issues, we design a mental disorder screening network, MDSNet, to explore a refined and tailored speech representation for SMS tasks, which consists of two components: 1) a multiscale temporal-frequency (MTF) encoder for capturing speech representations that are more adapted to various SMS tasks; and 2) a speech representation self-supervised disentanglement (S$^{3}$D) strategy for guiding the MTF encoder to disentangle the emotion-prosody representation from the unlabeled samples. To prove the effectiveness of the method, we validate it on a private dataset (AMHS) and two public datasets. The experimental results on the AMHS dataset demonstrate that our method can achieve promising performance across multiple mental disorder screening tasks. Additionally, our method achieves F1 scores of 0.850 and 0.987 on the NRAC and the CMDC datasets, respectively, outperforming the state-of-the-art methods. Taken together, our method provides a promising approach to mental disorder screening. https://github.com/1328915307/MDSNet.git.