ToxiAlert-Bench: leaderboard
Speech toxicity classification from audio (ToxiAlert-Bench, Ba et al., 2026): each English clip, real or synthesized, is safe or belongs to one of seven toxicity categories (sarcasm, horror, sexual, mental health and risk, ideology, violence and harm, discrimination), including toxicity carried only by tone or other paralinguistic cues. Macro-F1 in percent over the eight classes on the 6,519-clip test split, for audio language models prompted without task training; higher is better.
Metric: Macro-F1 (%; eight-way classification of each test clip as safe or one of seven toxicity categories, 6,519-clip test split; general-purpose audio LLM prompted zero-shot). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 57.55 | #452 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=toxialert-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.