KoALa-Bench - PA-QA: leaderboard
KoALa-Bench (2026) evaluates large audio language models on Korean speech understanding and on faithfulness to the speech context, under clean and noisy audio. This board is accuracy in percent on long spoken stories with multiple-choice questions whose evidence sits at a known position, clean audio; higher is better.
Metric: Accuracy (%) on PA-QA, MCTest stories read as long Korean speech with multiple-choice questions whose evidence sits at a known position, all positions, Korean speech synthesized or recorded per item, clean audio (first value of each clean / noisy pair), mean over four prompts with greedy decoding; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 92.51 | #928 |
| 2 | Gemini 2.5 Flash Lite | 92.31 | #972 |
| 3 | Voxtral-Mini-3B-2507 | 84.25 | #3545 |
| 4 | GPT-Audio Mini | 77.23 | #919 |
| 5 | Gemma 3n E4B (IT) | 48.7 | #2368 |
| 6 | Raon-Speech-9B | 40.98 |
No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=koala-bench-pa-qa · How It Works · Data refreshed daily, snapshot 2026-11-22.