KOpenAudioBench: leaderboard
Korean spoken question answering from KOpenAudioBench (Kim et al., 2026): 2,835 items transferred from OpenAudioBench, short-answer questions from LLaMA Questions, TriviaQA and WebQuestions plus open-ended AlpacaEval prompts, as TTS speech. Mean of the four sub-benchmark scores on a 0-100 scale (accuracy and GPT-5.4 judge score); higher is better.
Metric: Mean of the four sub-benchmark scores (short-answer accuracy on KLlamaQ, KTriviaQA and KWebQ, GPT-5.4 judge score on KAlpacaEval; 2,835 Korean spoken samples transferred from OpenAudioBench; 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 | 52.1 | |
| 2 | HyperCLOVA X 8B Omni | 45.1 | |
| 3 | Fun-Audio-Chat | 43.1 | |
| 4 | Qwen2.5-Omni | 39.2 | #1499 |
| 5 | MiniCPM-o 4.5 | 35.7 | |
| 6 | Step-Audio-2-mini | 31 | |
| 7 | Audio Flamingo 3 | 12.6 | |
| 8 | InteractiveOmni | 11.5 |
Interactive version: theaggregate.ai/benchmark?slug=kopenaudiobench · How It Works · Data refreshed daily, snapshot 2026-11-12.