KVoiceBench - KBBH: leaderboard
Spoken Korean binary reasoning questions transferred from BIG-Bench Hard, with Korean grammar judgments replacing adjective ordering, part of KVoiceBench (Kim et al., 2026). Accuracy in percent over 750 TTS-synthesized items; higher is better.
Metric: Accuracy (%; binary reasoning questions transferred from BIG-Bench Hard, with Korean grammar items in place of adjective ordering; Korean spoken questions synthesized with TTS; 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 | 83.5 | |
| 2 | MiniCPM-o 4.5 | 59.1 | |
| 3 | HyperCLOVA X 8B Omni | 52.7 | |
| 4 | Fun-Audio-Chat | 52.3 | |
| 5 | Qwen2.5-Omni | 49.9 | #1499 |
| 6 | Step-Audio-2-mini | 40.3 | |
| 7 | Audio Flamingo 3 | 36.3 | |
| 8 | InteractiveOmni | 8.3 |
Interactive version: theaggregate.ai/benchmark?slug=kvoicebench-kbbh · How It Works · Data refreshed daily, snapshot 2026-11-12.