SEA-SpeechBench - Emotion Recognition (SEA Prompt): leaderboard
Emotion recognition over a closed nine-class label set in SEA-SpeechBench (Liao et al., 2026) from short clips in 11 Southeast Asian languages, with the instruction written in the native Southeast Asian language. Accuracy in percent, with free-form answers mapped to labels by a Gemma-3-27B-Instruct judge; higher is better.
Metric: Judge-based accuracy (%; closed nine-class emotion label from speech, free-form answers mapped to labels and judged by Gemma-3-27B-Instruct, per-dataset scores averaged; short clips of at most 30 s from curated public Southeast Asian speech corpora, up to 1,000 sampled items per dataset; instruction prompt in the native Southeast Asian language). Source: arxiv.org. Saturation forecast: Around 2035. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MERaLiON-2-10B | 20.34 | #1646 |
| 2 | MERaLiON-2-3B | 18.73 | #6258 |
| 3 | Gemini 2.5 Flash | 16.79 | #452 |
| 4 | Gemma 3n E4B (IT) | 13.93 | #2220 |
| 5 | Gemma 3n E2B (IT) | 13.17 | #3758 |
| 6 | Qwen3 Omni 30B A3B Instruct | 11.31 | #902 |
| 7 | Qwen2.5-Omni-7B | 10.45 | #1797 |
| 8 | Phi-4 Multimodal Instruct | 9.2 | #3914 |
| 9 | Voxtral-Mini-3B-2507 | 5.35 | #3702 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sea-speechbench-emotion-recognition-sea-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.