AEQ-Bench - Naturalness: leaderboard
Empathetic response generation by omni-modal models on AEQ-Bench (Luo et al., 2026): given a human speech utterance and a short textual context (1,885 instances from Meld, GigaSpeech and EmoV-DB), the model replies as a supportive conversational partner. A GPT-5-mini judge rates how human-like and conversational each text reply is on a 1-4 scale, from pure analysis (1) to human-like chatting (4), averaged over the three subsets; higher is better.
Metric: Naturalness (1-4; how human-like and conversational the text response is, from analysis only (1) to human-like chatting (4), rated by a GPT-5-mini judge with manual checking; mean over the Meld, GigaSpeech and EmoV-DB subsets, 1,885 instances). Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen2.5-Omni-7B | 3.89 | #1797 |
| 2 | Qwen3 Omni 30B A3B Instruct | 3.77 | #902 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aeq-bench-naturalness · How It Works · Data refreshed daily, snapshot 2026-11-12.