Joint Distribution-Aligned Dual-Sparse Linear Regression for Cross-Stimulus Speech-Based Depression Detection

Abstract

This letter addresses the challenge of cross-stimulus speech-based depression detection (SDD), where training (source) and testing (target) speech samples stem from different stimulus methods, such as interview responses and reading texts. This discrepancy may create a mismatch in feature distributions between the source and target speech samples, leading to a notable deterioration in the performance of existing SDD methods. To tackle this issue, we propose a novel domain adaptation approach called Joint Distribution-aligned Dual-sparse Linear Regression (JDDLR). The fundamental idea of JDDLR is straightforward: extending simple linear regression (LR) to a version that is both depression-discriminative and stimulus-invariant. To achieve this, we initially equip JDDLR with depression-discriminative capability by constructing a dual-sparse linear regression (DLR) model. Unlike conventional linear regression models, DLR employs a meticulous coarse-to-fine feature selection mechanism to seek the depression-discriminative features from the acoustic feature set used to describe speech signals. Subsequently, we introduce a regularization term, which borrows the idea of joint distribution adaptation, thereby giving rise to JDDLR. This regularization term serves to alleviate the incongruities in feature distributions between the selected high-quality features of source and target samples. To evaluate JDDLR, extensive cross-stimulus SDD experiments are conducted on the MODMA dataset. The results underscore the promising performance of JDDLR in effectively addressing cross-stimulus SDD challenges.