KVoiceBench - KCommonEval: leaderboard
Spoken Korean open-ended spoken requests transferred from CommonEval, part of KVoiceBench (Kim et al., 2026). GPT-5.4 judge score on a 0-100 scale over 194 TTS-synthesized items; higher is better.
Metric: GPT-5.4 judge score (0-100; open-ended spoken requests transferred from CommonEval; 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 | Qwen2.5-Omni | 58.8 | #1499 |
| 2 | Raon-Speech | 57.4 | |
| 3 | Fun-Audio-Chat | 56.8 | |
| 4 | HyperCLOVA X 8B Omni | 54.2 | |
| 5 | MiniCPM-o 4.5 | 43.6 | |
| 6 | Step-Audio-2-mini | 32.4 | |
| 7 | Audio Flamingo 3 | 27.2 | |
| 8 | InteractiveOmni | 22.8 |
Interactive version: theaggregate.ai/benchmark?slug=kvoicebench-kcommoneval · How It Works · Data refreshed daily, snapshot 2026-11-12.