KMMAU - Fact Extraction: leaderboard
A fact stated in the audio, a four-way choice, on native Korean speech in KMMAU (Kim et al., 2026). Accuracy in percent; higher is better.
Metric: Accuracy (%; four-choice question on a fact stated in the audio; 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 | Fun-Audio-Chat | 91.9 | |
| 2 | MiniCPM-o 4.5 | 88.9 | |
| 3 | Raon-Speech | 86.9 | |
| 4 | Qwen2.5-Omni | 81.8 | #1499 |
| 5 | Step-Audio-2-mini | 79.8 | |
| 6 | Audio Flamingo 3 | 59.6 | |
| 7 | HyperCLOVA X 8B Omni | 36.4 | |
| 8 | InteractiveOmni | 33.3 |
Interactive version: theaggregate.ai/benchmark?slug=kmmau-fact-extraction · How It Works · Data refreshed daily, snapshot 2026-11-12.