CBT-Audio (Audio + Transcript): leaderboard
Rating the distress intensity (1 to 5) of a patient turn in public role-play and case-study CBT session recordings, given the three preceding turns as text; CBT-Audio (Hu et al., 2026). The model gets both the turn's audio and its transcript. Mean absolute error in scale points against reference labels built from GPT-audio-1.5 descriptions and checked on an expert-annotated subset; lower is better.
Metric: Mean absolute error on the 1-5 distress-intensity scale against SSR reference labels (1,802 patient turns; lower is better); patient audio, its transcript and text context. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | gemma-4-E4B-it | 0.72 | #1462 |
| 2 | Qwen2.5-Omni-7B | 0.93 | #1797 |
| 3 | Voxtral-Mini-3B-2507 | 1.1 | #3702 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cbt-audio-audio-plus-transcript · How It Works · Data refreshed daily, snapshot 2026-11-12.