DEPOOL benchmark: 1/3 of depression detection models collapse to single class

A new benchmark study from researchers at USC (Anisha Pattanayak, Huang-Cheng Chou, Shrikanth Narayanan, Sudarsana Reddy Kadiri) takes a rigorous look at how speech-based depression detection models aggregate features from short audio segments into a speaker-level decision. The paper, submitted to SLT 2026, introduces DEPOOL — a framework that compares six temporal aggregation architectures across six frozen self-supervised speech backbones, using both an English and a Mandarin depression corpus. Instead of fixing a single encoder layer by hand, DEPOOL learns which backbone layers matter for each configuration, creating a grid of 72 distinct pipelines.

The results are sobering. One-third of all configurations collapsed into predicting a single class for every speaker — a failure mode that was tied both to the aggregation method and the backbone choice. Furthermore, architectures that appeared stable in a single random seed often became unreliable when training was repeated across seeds. The authors argue that robustness to backbone and seed variation should be a first-class criterion for benchmarking temporal aggregation, rather than relying on average accuracy from a single pipeline. This work has direct implications for deploying AI in clinical mental health assessments, where model reliability is paramount.