SEA-SpeechBench - Timestamped Content Query (60-120 s): leaderboard
Metric: WER or CER (%; transcribe only the speech inside a queried time window of a long recording of 60-120 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: Around October 2026. 10 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 1.34 |
| 2 | GPT-4o Audio | 7.38 |
| 3 | Gemini 2.5 Flash | 9.44 |
| 4 | Qwen2.5-Omni-7B | 9.85 |
| 5 | Voxtral-Mini-3B-2507 | 12.86 |
| 6 | Phi-4 Multimodal Instruct | 16.25 |
Interactive version: theaggregate.ai/benchmark?slug=sea-speechbench-timestamped-content-query-60-120-s · How It Works · Data refreshed daily, snapshot 2026-09-26.