SEA-SpeechBench - Temporal Localization (30-60 s): leaderboard
Temporal localization in SEA-SpeechBench (Liao et al., 2026): give the start and end time at which a queried phrase is spoken in a Southeast Asian recording lasting 30 to 60 seconds, with an English instruction prompt. Span-overlap F1 in percent; higher is better.
Metric: Span-overlap F1 (%; predict the start and end time at which a queried phrase is spoken in a recording of 30-60 seconds; English instruction prompt; blank where the model cannot take audio of that length). Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen2.5-Omni-7B | 19.98 | #1797 |
| 2 | MERaLiON-2-10B | 12.37 | #1646 |
| 3 | MERaLiON-2-3B | 10.28 | #6258 |
| 4 | Voxtral-Mini-3B-2507 | 9.87 | #3702 |
| 5 | Gemma 3n E4B (IT) | 8.72 | #2220 |
| 6 | Gemma 3n E2B (IT) | 8.27 | #3758 |
| 7 | Gemini 2.5 Flash | 7.7 | #452 |
| 8 | Phi-4 Multimodal Instruct | 6.25 | #3914 |
| 9 | Qwen3 Omni 30B A3B Instruct | 5.9 | #902 |
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-temporal-localization-30-60-s · How It Works · Data refreshed daily, snapshot 2026-11-12.