MCGA - ASR - Fu: leaderboard
Speech recognition of 200 recorded fu rhapsodies read by native Mandarin speakers (MCGA test set; Du et al., Harbin Institute of Technology, 2026). Character error rate in percent against the original text; lower is better. The authors' Qwen-Omni-MCGA row is a LoRA fine-tune of Qwen2.5-Omni-7B on the MCGA training split.
Metric: Character error rate (%; 200 recorded fu rhapsodies of the test set). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 6.2 | #902 |
| 2 | Gemini 3 Flash | 8.4 | #159 |
| 3 | Qwen2.5-Omni-7B | 14.8 | #1797 |
| 4 | Voxtral-Small-24B-2507 | 33.9 | #1291 |
| 5 | Voxtral-Mini-3B-2507 | 34.5 | #3702 |
| 6 | Phi-4 Multimodal Instruct | 61.1 | #3914 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mcga-asr-fu · How It Works · Data refreshed daily, snapshot 2026-11-12.