ParaPairAudioBench - Style: leaderboard
ParaPairAudioBench (Jeon et al., Interspeech 2026) tests audio language models as pairwise judges: given two recordings and a speech-generation instruction, the model picks the one that fulfils it better or declares a tie, and accuracy in percent is averaged over both presentation orders; higher is better. This board is 1,440 Expresso pairs judged on a requested speaking style (whisper, happy, sad, confused and others), with ties allowed, so chance is one third.
Metric: Pairwise judgment accuracy (%; mean of both A/B orders; A better, B better or tie; 1,440 Expresso pairs, target speaking style). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 48.5 | #452 |
| 2 | Qwen2.5-Omni-7B | 35.8 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=parapairaudiobench-style · How It Works · Data refreshed daily, snapshot 2026-11-12.