ODU-Bench (Audio-Visual) - Transcription: leaderboard
Transcription in ODU-Bench (Chen et al., 2026): accuracy of the transcript of the spoken demand, one minus the pooled character or word error rate, on the demand scenes of 1,347 audio-visual scenes (1,070 synthetic and 277 human-recorded; user turns as video with audio, earlier assistant replies as text). Percent; higher is better.
Metric: Transcript accuracy (%; max(0, 1 - pooled CER or WER) of the demand transcript, demand scenes only; 1,347 audio-visual scenes (1,061 with a demand, 286 without; 1,070 synthetic and 277 human-recorded), user turns as video with audio, earlier assistant replies as text). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 91.3 | #91 |
| 2 | Qwen3 Omni 30B A3B Instruct | 87.1 | #902 |
| 3 | Gemini 3.7 Flash | 85.9 | #32 |
| 4 | Qwen 3.5 Omni Plus | 83.3 | #237 |
| 5 | Seed 2.0 Lite | 77.7 | #286 |
| 6 | Qwen2.5-Omni-7B | 74.2 | #1797 |
| 7 | Gemini 3.5 Flash Lite | 71.5 | #300 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=odu-bench-audio-visual-transcription · How It Works · Data refreshed daily, snapshot 2026-11-12.