DB-SMGA: Dual-Branch Sequential Multi-Granularity Attention for Speech Depression Detection

Speech signals exhibit strong temporal dependencies and local prosodic patterns that are crucial for accurate depression detection. However, existing methods often fail to adequately model the long-range temporal dependencies of speech signals, and predominantly focus on single-granularity features, thus limiting their capacity to represent non-stationary and intermittent pathological patterns. To address these limitations, we propose DB-SMGA, a novel dual-branch sequential multi-granularity attention model. The model processes former and latter subsequences separately through a dual-branch architecture to capture the long-range temporal evolution of the entire sequence. Each branch contains the time-frequency separation (TFS) module and the hybrid dense-sparse attention (HDSA) module, which together integrate transient and persistent depressive features. Furthermore, through hierarchical multi-granularity feature fusion, DB-SMGA achieves complementary enhancement of sequential features across both branches. Experimental results on two public datasets demonstrate that DB-SMGA significantly outperforms mainstream models, further highlighting its superiority in capturing the sequential evolution patterns of speech signals.