KVoiceBench - KAdvBench: leaderboard
Safety on spoken Korean harmful requests transferred from AdvBench, part of KVoiceBench (Kim et al., 2026). Refusal rate in percent over 513 TTS-synthesized requests, detected by rules; higher is better.
Metric: Refusal rate (%; rule-based detection of refusals to harmful spoken requests transferred from AdvBench; Korean spoken questions synthesized with TTS; single run at temperature 0.7). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen2.5-Omni | 95.9 | #1499 |
| 2 | Raon-Speech | 87.3 | |
| 3 | Fun-Audio-Chat | 84.8 | |
| 4 | HyperCLOVA X 8B Omni | 83.2 | |
| 5 | InteractiveOmni | 68 | |
| 6 | Step-Audio-2-mini | 64.3 | |
| 7 | MiniCPM-o 4.5 | 48.5 | |
| 8 | Audio Flamingo 3 | 5.7 |
Interactive version: theaggregate.ai/benchmark?slug=kvoicebench-kadvbench · How It Works · Data refreshed daily, snapshot 2026-11-12.