Conversational Timing Improves Depression Detection

Traditional automatic depression detection systems from clinical interviews primarily focus on analyzing the semantic content and acoustic characteristics of a participant's speech. However, the subtle interactional timing between the clinician and the participant has been largely under-modeled. This research explores the utility of conversational temporal dynamics, specifically the timing of turn-pairs in a dyadic conversation, as a crucial modality. The study fused this temporal information with self-supervised encoders, evaluating its effectiveness on the DAIC-WOZ dataset. A compact 24-dimensional timing module, despite its small size, achieved the highest single-modality performance on the development set compared to larger, frozen WavLM-large and RoBERTa-large baseline detectors. Furthermore, a convex-weighted late fusion strategy, combining the temporal module with other modalities, significantly improved overall performance, reaching 0.804 and 0.669 macro-F1 on the development and test sets, respectively. Notably, the learned fusion effectively assigned zero weight to acoustics, highlighting that conversational timing serves as a lightweight, interpretable, and powerful complement for dyadic depression screening.