VocalAffectBench: leaderboard
VocalAffectBench (2026) is a public, test-only benchmark of 280 human-recorded English speech clips, 40 for each of seven emotions performed from scripts that fit several emotions, testing whether audio models recognize the emotion expressed in the delivery rather than in the words. Seven-way accuracy in percent from raw audio alone; higher is better. Prompted audio models get a fixed label list, and provider emotion endpoints have their native labels mapped to the seven classes.
Metric: Accuracy (%; seven-way expressed vocal emotion classification (angry, disgusted, fearful, happy, neutral, sad, surprised) of 280 human-recorded English clips, 40 per label, from raw audio alone with no transcript; prompted audio models get a fixed instruction listing the labels, provider emotion or prosody endpoints run at default settings, and native outputs are mapped to the label set before scoring). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash | 44.3 | #98 |
| 2 | Qwen 3.5 Omni Plus | 37.9 | #237 |
| 3 | Voxtral-Small-24B-2507 | 34.3 | #1291 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vocalaffectbench · How It Works · Data refreshed daily, snapshot 2026-11-12.