The pauses and rhythms between speakers in a clinical interview may carry diagnostic signal that most AI depression-screening tools quietly ignore.
Researchers tested whether dyadic turn-pair timing — specifically, how long each party waits before responding — could work as a primary detection signal rather than an afterthought. Evaluated on the DAIC-WOZ benchmark dataset, a compact 24-dimensional timing module outperformed frozen WavLM-large and RoBERTa-large models as a single modality on the development set. When the team fused the timing module with those larger models using a convex-weighted late fusion strategy, the combined system reached macro-F1 scores of 0.804 on development and 0.669 on test sets. Notably, the learned fusion assigned zero weight to acoustics — the voice-tone features that most existing systems treat as essential.
Most depression-screening AI focuses on what people say or how they sound saying it; this work suggests that when they say it matters too. A lightweight timing module that complements a language model without needing the acoustic layer could lower the computational cost of deployment and reduce one class of potential bias introduced by voice-quality features.
The study is preliminary and evaluated on a single dataset, so generalization to real clinical settings is still an open question — which is the kind of caveat that tends to get lost between a preprint and a press release.