VocalGrad - Roughness: leaderboard
VocalGrad (2026), 21,600 binary acoustic-change questions across nine transformations; case-insensitive whole-word direction parsing with exactly one direction required; invalid outputs fail. Gemini 3 Flash uses low thinking and temperature 1; MiMo temperature 0.3/top-p 0.95; other open models greedy where available. Main model runs use the complete dataset, not the differently sampled six-person human study. Accuracy (%); higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Step-Audio 2 mini | 49.9 | |
| 2 | Kimi-Audio | 48 | |
| 3 | Audio Flamingo 3 | 46.7 | |
| 4 | Gemini 3 Flash (Low) | 44.8 | #99 (Gemini 3 Flash) |
| 5 | MiMo-Audio | 43.2 |
Interactive version: theaggregate.ai/benchmark?slug=vocalgrad-roughness · How It Works · Data refreshed daily, snapshot 2026-10-20.