STEB - Emotion Preservation (English to Chinese): leaderboard
English to Chinese speech-to-speech translation on STEB, a 32.6-hour benchmark of real drama, audiobook, advertisement, interview, news and commentary speech (Cheng et al., 2026). Emotion consistency between source and translated speech on a 1-5 scale, judged by Qwen3-30B-A3B from Qwen3-Omni-Captioner descriptions of both; higher is better.
Metric: LLM-judged emotion consistency (1-5) of category, intensity and attitude between the source and the translated speech: both are captioned by Qwen3-Omni-Captioner and summarized by Qwen3-30B-A3B, which scores the pair three times (outlier-aware aggregate); English-to-Chinese speech-to-speech translation of the normal subset's English source utterances. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | UniSS | 3.82 | |
| 2 | Step-Audio-2 | 3.77 | |
| 3 | Seed LiveInterpret 2.0 | 3.53 | |
| 4 | SeamlessExpressive | 3.27 |
Interactive version: theaggregate.ai/benchmark?slug=steb-emotion-preservation-english-to-chinese · How It Works · Data refreshed daily, snapshot 2026-11-12.