VoxSafeBench - Symbolic Background (Chinese): leaderboard
VoxSafeBench (2026) Tier 2 safety test for speech language models with Chinese audio where the words are harmless but the audio is not: the background holds harmful audio such as extremist anthems that the user asks to amplify. The score is the percentage of responses that warn or refuse while naming the audio cue, judged by DeepSeek-V3, higher is better.
Metric: Safety awareness rate (%), the last of the four DAR / WAR / RtA / SAR parts: VoxSafeBench Tier 2 spoken requests that are safe for a neutral adult but call for caution because of an audio cue; DeepSeek-V3 labels each response as a direct answer, a warning that names the cue, a refusal that names the cue, or skip, and the score is warnings plus refusals; Chinese audio; here the background holds harmful audio such as extremist anthems that the user asks to amplify; higher is better. Source: arxiv.org. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemini 3 Pro | 54.4 | #145 | |
| 2 | Gemini 3 Flash | 36.9 | #175 | |
| 3 | Qwen 3 Omni 30B A3B (Thinking) | 13.1 | ||
| 4 | Qwen3 Omni 30B A3B Instruct | 6.9 | #930 | |
| 5 | MiMo-Audio-7B-Instruct | Thinking | 6.9 | #1336 |
| 6 | Kimi-Audio | 6.2 | #1735 | |
| 7 | GPT-4o Audio | 2.5 | #1129 |
Interactive version: theaggregate.ai/benchmark?slug=voxsafebench-symbolic-background-chinese · How It Works · Data refreshed daily, snapshot 2026-11-21.