MCGA - Speech Emotion Captioning: leaderboard
Describing the speaker and the emotions of a recited classical Chinese text from audio (MCGA; Du et al., Harbin Institute of Technology, 2026), 1,000 test recordings. Emotion Caption Fidelity, judged by DeepSeek-V3.2, sums persona recognition (20 points), global emotional tone (30) and sentence-level emotion (50); higher is better.
Metric: Emotion Caption Fidelity (0-100; persona, global tone and sentence-level emotion of 1,000 test recordings, DeepSeek-V3.2 judge). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 58.4 | #902 |
| 2 | Gemini 3 Flash | 54 | #159 |
| 3 | Qwen2.5-Omni-7B | 37 | #1797 |
| 4 | Voxtral-Small-24B-2507 | 16.4 | #1291 |
| 5 | Voxtral-Mini-3B-2507 | 15 | #3702 |
| 6 | Phi-4 Multimodal Instruct | 12.7 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mcga-speech-emotion-captioning · How It Works · Data refreshed daily, snapshot 2026-11-12.