Conversational Timing improves depression detection on DAIC-WOZ

Hanie Kang, Huang-Cheng Chou, Sudarsana Reddy Kadiri and Shrikanth Narayanan submitted an arXiv paper on 4 Jul 2026 that tests whether conversational temporal dynamics can improve automatic depression detection in dyadic interviews. They evaluate a compact 24-dimensional timing module on the DAIC-WOZ dataset and compare it to frozen WavLM-large and RoBERTa-large baselines.

What did the authors test?

They investigated conversational temporal dynamics, specifically dyadic turn-pair timing, as a primary modality fused with self-supervised encoders, and evaluated the approach on DAIC-WOZ. The timing module is 24-dimensional; it was compared against frozen WavLM-large and RoBERTa-large baseline detectors and then combined with them using a convex-weighted late fusion strategy.

The experiments contrast single-modality detectors with multimodal fusion. The paper frames timing as a lightweight modality and pairs it with larger, frozen encoders rather than training end-to-end models from scratch.

How well did conversational timing perform?

The 24-dimensional temporal module achieved the highest single-modality performance on the development set, and a convex-weighted late fusion improved overall performance to 0.804 macro-F1 on the development set and 0.669 macro-F1 on the test set. The learned fusion assigned zero weight to acoustics, indicating the timing signal dominated the combined model.

Those are the concrete numbers reported: a development macro-F1 of 0.804 and a test macro-F1 of 0.669 for the convex-weighted late fusion. The paper highlights that timing alone beat each frozen baseline on the development split, though the authors do not print exact single-modality numeric scores for every baseline in the abstract.

Why it matters

Timing provides a compact, interpretable signal for dyadic depression screening, the authors argue. A 24-dimensional module that beats large frozen encoders on development data suggests conversational structure contains clinically relevant cues beyond what semantic and acoustic encoders capture. The fusion result and the observation that learned weights set acoustic contribution to zero underscore that timing can act as a lightweight complement to heavyweight encoders.

This matters for applied screening systems because a small timing encoder is cheaper to store and faster to run than large self-supervised models, and its interpretability can aid clinical review. The paper was submitted to SLT 2026, positioning the work for peer review and community scrutiny.

What to watch

Track the paper through SLT 2026, where it was submitted, and look for fuller results or a published version that reports per-modality numeric breakdowns beyond the abstract summary. Replication on datasets beyond DAIC-WOZ and any released code or model checkpoints would be the next concrete signals that the approach generalizes.

Methods and data summary

• Dataset: DAIC-WOZ, used for dyadic depression detection tasks.
• Timing module: compact, 24-dimensional dyadic turn-pair timing features.
• Baselines: frozen WavLM-large and RoBERTa-large detectors.
• Fusion: convex-weighted late fusion, which reached 0.804 dev macro-F1 and 0.669 test macro-F1; fusion learned to assign zero weight to acoustics.

Authors and provenance The paper lists Hanie Kang, Huang-Cheng Chou, Sudarsana Reddy Kadiri and Shrikanth Narayanan and was submitted to arXiv on 4 Jul 2026, with a submission note that it was submitted to SLT 2026.

For readers: the abstract frames conversational timing as a lightweight, interpretable complement for dyadic depression screening and reports the development and test macro-F1 numbers above as the main quantitative outcomes.