SpeechParaling-Bench (Chinese) - Paralanguage Control: leaderboard
SpeechParaling-Bench (2026) tests paralinguistic-aware speech generation by audio-language models; a Gemini 3 Pro judge compares each spoken response with the response of a fixed baseline model, ties counting half, and the baseline itself scores its strength-weighted result against the others. This board covers the paralanguage control task (repeat a sentence with specified vocal features or style) in the Chinese subset, against the Doubao Realtime Voice baseline. Pairwise score in percent; higher is better.
Metric: Pairwise score (%) on the paralanguage control task (repeat a sentence with specified vocal features or an abstract style), Chinese subset: a Gemini 3 Pro judge compares each spoken response with the fixed baseline's (Doubao Realtime Voice; ties count half), and the baseline's own row is its strength-weighted average against the candidates (the paper's Eq. 3); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Doubao Realtime Voice Model | 71.86 | |
| 2 | GPT-Audio (2025-08-28) | 35.57 | |
| 3 | Gemini 2.5 Flash Audio | 29.64 | |
| 4 | Qwen3-Omni-Flash-2025-12-01 | 14.16 | |
| 5 | Qwen3 Omni Realtime | 4.72 |
Interactive version: theaggregate.ai/benchmark?slug=speechparaling-bench-chinese-paralanguage-control · How It Works · Data refreshed daily, snapshot 2026-11-21.