CASU - Role Inference: leaderboard
Metric: Accuracy (%; four-option multiple-choice questions on semi-synthetic auditory scenes that combine speech, sound events and background; role inference: infer a speaker's role or the relationship between two speakers from speech and scene). Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.0 Flash | 69.05 |
| 2 | GPT-4o Audio | 68.58 |
| 3 | Qwen3 Omni 30B A3B Instruct | 63.51 |
| 4 | Voxtral-Small-24B-2507 | 51.64 |
| 5 | Qwen2.5-Omni-7B | 51.37 |
| 6 | Qwen2-Audio-7B-Instruct | 40 |
| 7 | SALMONN-13B | 39.73 |
Interactive version: theaggregate.ai/benchmark?slug=casu-role-inference · How It Works · Data refreshed daily, snapshot 2026-09-26.