SEA-SpeechBench - Spoken QA (English Prompt): leaderboard
Spoken question answering in SEA-SpeechBench (Liao et al., 2026) over long-form Southeast Asian speech from YODAS2, with an English instruction prompt. A GPT-4.1 judge rates each short answer from 0 to 5, scaled to 0 to 100; higher is better.
Metric: Judge score (0-100): short answers to questions about long-form Southeast Asian speech from YODAS2, rated 0-5 by a GPT-4.1 judge and multiplied by 20; English instruction prompt. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 92.28 | #452 |
| 2 | Qwen3 Omni 30B A3B Instruct | 85.77 | #902 |
| 3 | Voxtral-Mini-3B-2507 | 82.32 | #3702 |
| 4 | MERaLiON-2-10B | 82 | #1646 |
| 5 | Gemma 3n E4B (IT) | 79.24 | #2220 |
| 6 | Gemma 3n E2B (IT) | 73.35 | #3758 |
| 7 | Qwen2.5-Omni-7B | 71.6 | #1797 |
| 8 | MERaLiON-2-3B | 66.68 | #6258 |
| 9 | Phi-4 Multimodal Instruct | 64.69 | #3914 |
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-spoken-qa-english-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.