KVoiceBench - KIFEval: leaderboard
Spoken Korean instruction-following prompts transferred from IFEval, part of KVoiceBench (Kim et al., 2026). Mean of prompt-level and instruction-level accuracy in percent over 279 TTS-synthesized items; higher is better.
Metric: Instruction-following accuracy (%; mean of prompt-level and instruction-level accuracy on instructions transferred from IFEval; 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 | Raon-Speech | 71.6 | |
| 2 | Fun-Audio-Chat | 53.4 | |
| 3 | Qwen2.5-Omni | 42.9 | #1499 |
| 4 | MiniCPM-o 4.5 | 34.6 | |
| 5 | HyperCLOVA X 8B Omni | 32.6 | |
| 6 | Step-Audio-2-mini | 30.4 | |
| 7 | Audio Flamingo 3 | 20.8 | |
| 8 | InteractiveOmni | 10.7 |
Interactive version: theaggregate.ai/benchmark?slug=kvoicebench-kifeval · How It Works · Data refreshed daily, snapshot 2026-11-12.