Paper Exposes Inflated Evaluations in Speech-Based Depression Detection

A new paper from TU München researchers shows speech-based depression detection evaluations suffer from selection bias, with layer-picking inflating AUC by 0.07 on average, and label-shuffled test sets still yielding AUC 0.59 via best-of-25 selection.

2026-10-16 ~ 2026-10-16 · 2 related posts

• Shuffled labels hit AUC 0.59: selection bias, not signal, drives prosody benchmarks — maier_ak · 2026-10-16
• Speech depression detection AUC inflated by 0.07 via layer selection, preprint shows — maier_ak · 2026-10-16