MSU-Bench: leaderboard
MSU-Bench (2026) tests speaker-centric understanding in conversational audio with several speakers: 2,300 human-verified four-option questions over 16 tasks, from identifying who spoke and their attributes to reasoning about roles, dialogue structure and emotions. The score is instance-level accuracy over all tasks, in percent; higher is better.
Metric: Overall accuracy (%), instance-level over all 16 tasks, four-option multiple choice on MSU-Bench (2,300 human-verified questions over Chinese and English multi-speaker conversations from telephone, meeting, podcast and movie audio), zero-shot with one instruction template requiring a single option letter, exact-match accuracy (x100); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 77 | #164 |
| 2 | Gemini 2.5 Pro | 70 | #266 |
| 3 | Gemini 2.5 Flash | 69 | #453 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=msu-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.