Attention-Guided Bi-direction Temporal-aware Network for speech-based depression recognition

Depression is a serious mental illness that affects daily life and has drawn increasing global concern. While speech contains valuable emotional markers for depression recognition, accurate estimation remains challenging. We propose the Attention Guided Bi-direction Temporal-aware Network (AGBiTNet), a novel architecture designed for speech-based depression recognition. AGBiTNet incorporates a Bi-direction Temporal-aware Module (BiTM) to capture bidirectional temporal dependencies and a Frequency-aware Attention Module (FAM) to extract discriminative emotional cues from multi-scale spectral representations. To further enhance regression accuracy and improve feature robustness, a joint loss combining Huber loss and Generalized End-to-End (GE2E) loss is adopted. Extensive experiments on AVEC 2013, 2014, and 2017 datasets demonstrate that AGBiTNet achieves competitive performance with RMSE/MAE values of 9.36/7.21, 9.38/7.24, and 5.29/4.20, respectively. Ablation and statistical analyses confirm the reliability of these results, highlighting the proposed approach as an effective and lightweight solution for speech-based depression assessment with promising practical applicability.