SonicBench - Counting: leaderboard
SonicBench (Sun et al., 2026) tests whether audio language models perceive physical properties of sound with two-option questions built from controlled contrasts far above human discrimination thresholds, each attribute asked as recognition of one clip and as comparison of two clips; abstentions count as wrong and random guessing scores 50. This board is the 200 questions on the number of sound events in a scene (PicoAudio); accuracy in percent, higher is better.
Metric: Accuracy (%; number of sound events (PicoAudio); 100 recognition and 100 comparison two-option questions, greedy decoding, abstentions scored wrong; random 0.50). Source: arxiv.org. Saturation forecast: Estimated already saturated. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 82 | #902 |
| 2 | Gemini 2.5 Flash | 71 | #452 |
| 3 | Qwen2.5-Omni-7B | 58 | #1797 |
| 4 | Voxtral-Mini-3B-2507 | 51 | #3702 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sonicbench-counting · How It Works · Data refreshed daily, snapshot 2026-11-12.