Dataset domain invariant multitask learning for speech based depression detection

Abstract

Depressive disorder is a frequent and debilitating mental illness that must be identified early and monitored on an ongoing basis to be managed. The emergence of speech as a potential non-invasive behavioural biomarker for depression assessment is related to the presence of prosodic, temporal and voice-quality features of depression that can be detected. Current depression detection methods based on speech, however, rely on handcrafted acoustic features and/or black-box deep learning models, have limited usability across different datasets, and are mostly limited to binary screening rather than clinically relevant severity estimation. To overcome these drawbacks, this paper introduces the DepressSenseAI framework and the proposed VoxDepressNet model for effective voice-based depression assessment. It adopts a hybrid fusion strategy that combines self-supervised speech representations with interpretable acoustic and prosodic features. Attention-guided segment-to-session aggregation is used to preserve long-range temporal dependencies and highlight speech segments useful for diagnosis. In addition, a multi-task learning (MTL) approach is used to jointly learn the depression classification and severity estimation tasks when clinicians provide severity annotations, and a domain-adversarial learning (DAL) approach is applied to reduce dataset-specific information in the shared representation and improve robustness across heterogeneous speech corpora. Detailed experiments were conducted on a variety of public speech databases, including ADC, EATD, and PDCH. VoxDepressNet outperformed strong deep learning baselines, achieving up to 0.86 UAR for depression classification and 2.85 MAE for estimating depression severity from clinician-rated HAMD-17 scores. Cross-dataset experiments across diverse language and recording conditions confirmed this. The outcomes indicate the feasibility of using DepressSenseAI for scalable depression screening and mental health assessment of depression severity. The implementation code and experimental configurations are publicly available at https://github.com/mnaveed07154/DepressSenseAI. Not applicable.