VoxParadox: leaderboard
Paralinguistic listening under contradiction (VoxParadox, Pang, Chaubey and Soleymani, 2026): 2,000 multiple-choice questions over synthesized speech in ten tasks (age, gender, emotion, pitch, volume, speed, vocal range, intonation, speaker identity and speaker counting), where the spoken words always claim a wrong answer, such as several voices saying only one person is speaking. The score is accuracy against the acoustic ground truth in percent, averaged over the ten tasks; higher is better. Models that follow the transcript score below random guessing.
Metric: Ground-truth accuracy (%; mean over ten paralinguistic tasks of 200 multiple-choice questions each, transcripts assert a wrong answer). Source: arxiv.org. Saturation forecast: Around July 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 24.7 | #452 |
| 2 | Qwen2.5-Omni-7B | 7.95 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=voxparadox · How It Works · Data refreshed daily, snapshot 2026-11-12.