SEA-SpeechBench - Timestamped Content Query (0-30 s): leaderboard

Metric: WER or CER (%; transcribe only the speech inside a queried time window of a long recording of 0-30 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. 15 models tracked.

Top models

#ModelScore
1Qwen3 Omni 30B A3B Instruct1.18
2Gemini 2.5 Flash2.41
3Voxtral-Mini-3B-25074.74
4GPT-4o Audio5.06
5SeaLLMs-Audio-7B5.48
6Qwen2.5-Omni-7B5.49
7Qwen2-Audio-7B-Instruct5.56
8Gemma 3n E4B (IT)6.76
9gemma-3n-E2B-it7.32
10Phi-4 Multimodal Instruct20.67

Interactive version: theaggregate.ai/benchmark?slug=sea-speechbench-timestamped-content-query-0-30-s · How It Works · Data refreshed daily, snapshot 2026-09-26.