STEB - Nonverbal Vocalization Preservation (Chinese to English): leaderboard
Chinese to English 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). Preservation of nonverbal vocalizations such as laughter (type, count, order and position) on the nonverbal-vocalization subset, on a 1-5 scale judged by Qwen3-30B-A3B; higher is better.
Metric: LLM-judged nonverbal-vocalization preservation (1-5): type, count, order and position of events such as laughter in the NV-marked transcripts of the source and the translated speech (BEATs event detection), scored three times by Qwen3-30B-A3B (outlier-aware aggregate); Chinese-to-English speech-to-speech translation of the NV subset's Chinese source utterances. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Step-Audio-2 | 1.58 | |
| 2 | UniSS | 1.31 | |
| 3 | SeamlessExpressive | 1.29 | |
| 4 | Seed LiveInterpret 2.0 | 1.25 |
Interactive version: theaggregate.ai/benchmark?slug=steb-nonverbal-vocalization-preservation-chinese-to-english · How It Works · Data refreshed daily, snapshot 2026-11-12.