KMMAU: leaderboard
Korean audio understanding from KMMAU (Kim et al., 2026): 2,204 multiple-choice questions on native Korean speech from KSS, KMSAV and the Seoul Corpus about speaker attributes and spoken content. Mean accuracy in percent over its nine capabilities; higher is better.
Metric: Accuracy (%; mean of the nine capability accuracies over 2,204 multiple-choice questions on native Korean audio from KSS, KMSAV and the Seoul Corpus; single run at temperature 0.7). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Raon-Speech | 71.8 | |
| 2 | Fun-Audio-Chat | 70.4 | |
| 3 | Step-Audio-2-mini | 64.8 | |
| 4 | Qwen2.5-Omni | 62.9 | #1499 |
| 5 | MiniCPM-o 4.5 | 62.9 | |
| 6 | Audio Flamingo 3 | 45.2 | |
| 7 | HyperCLOVA X 8B Omni | 31.6 | |
| 8 | InteractiveOmni | 30.7 |
Interactive version: theaggregate.ai/benchmark?slug=kmmau · How It Works · Data refreshed daily, snapshot 2026-11-12.