KoALa-Bench - SIF Alpaca: 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 the GPT-4o-judged instruction-following score, on a 0 to 100 scale, for spoken Korean Alpaca instructions, clean audio; higher is better.

Metric: Speech instruction-following score (GPT-4o judge returning 0-1 against the reference answer, x100) on the Korean spoken Alpaca instructions, 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

#ModelScoreOverall rank
1GPT-Audio Mini90.58#918
2Gemini 2.5 Flash Lite85.94#964
3Gemma 3n E4B (IT)82.83#2401
4Qwen3 Omni 30B A3B Instruct82.55#930
5Raon-Speech-9B76.34
6Voxtral-Mini-3B-250768.48#3682

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=koala-bench-sif-alpaca · How It Works · Data refreshed daily, snapshot 2026-11-21.