KVoiceBench: leaderboard
Korean spoken question answering from KVoiceBench (Kim et al., 2026): 7,306 items transferred from VoiceBench with ground-truth correction, Korean redesign of language-specific items, speech-friendly normalization and TTS synthesis. Mean of nine sub-benchmark scores on a 0-100 scale (accuracy, GPT-5.4 judge scores, instruction-following accuracy and refusal rate); higher is better.
Metric: Mean of the nine sub-benchmark scores (each on 0-100 under its source metric: accuracy, GPT-5.4 judge score, instruction-following accuracy or refusal rate; 7,306 Korean spoken samples transferred from VoiceBench; 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 | 66.6 | |
| 2 | Fun-Audio-Chat | 50.1 | |
| 3 | Qwen2.5-Omni | 49 | #1499 |
| 4 | HyperCLOVA X 8B Omni | 45.1 | |
| 5 | MiniCPM-o 4.5 | 39.5 | |
| 6 | Step-Audio-2-mini | 32 | |
| 7 | InteractiveOmni | 20 | |
| 8 | Audio Flamingo 3 | 18.8 |
Interactive version: theaggregate.ai/benchmark?slug=kvoicebench · How It Works · Data refreshed daily, snapshot 2026-11-12.