SEA-SpeechBench - Timestamped Content Query (30-60 s): leaderboard
Timestamped content query in SEA-SpeechBench (Liao et al., 2026): transcribe only what is said inside a given time window of a Southeast Asian recording lasting 30 to 60 seconds, with an English instruction prompt. Word or character error rate in percent, which can exceed 100; lower is better.
Metric: WER or CER (%; transcribe only the speech inside a queried time window of a long recording of 30-60 seconds, CER for languages without word boundaries; English instruction prompt; blank where the model cannot take audio of that length; can exceed 100; lower is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 1.22 | #902 |
| 2 | Gemini 2.5 Flash | 2.64 | #452 |
| 3 | Qwen2.5-Omni-7B | 6.58 | #1797 |
| 4 | Gemma 3n E4B (IT) | 7.38 | #2220 |
| 5 | Gemma 3n E2B (IT) | 7.49 | #3758 |
| 6 | Voxtral-Mini-3B-2507 | 7.74 | #3702 |
| 7 | MERaLiON-2-3B | 8.21 | #6258 |
| 8 | MERaLiON-2-10B | 9.48 | #1646 |
| 9 | Phi-4 Multimodal Instruct | 14.9 | #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-timestamped-content-query-30-60-s · How It Works · Data refreshed daily, snapshot 2026-11-12.