HELM Audio - Mustard: leaderboard
The "Mustard" subtask of HELM Audio. Stanford HELM's audio project (AHELM): audio-language models are given speech and sound clips and scored on transcription, spoken-question answering, reasoning over audio, speaker and emotion attributes, safety and fairness.
Metric: EM. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.0 Flash Lite | 66.09 | #455 |
| 2 | Gemini 2.5 Pro (Preview 05-06) | 65.51 | #149 |
| 3 | Gemini 2.0 Flash | 60.43 | #345 |
| 4 | Gemini 2.0 Flash (Preview) | 60.43 | #369 |
| 5 | Qwen2.5-Omni-7B | 58.84 | #634 |
| 6 | Gemini 1.5 Pro (002) | 57.68 | #352 |
| 7 | Gemini 2.5 Flash (Preview 05-20) | 57.39 | #217 |
| 8 | Gemini 1.5 Pro (001) | 56.38 | #461 |
| 9 | Gemini 1.5 Flash (002) | 55.8 | #584 |
| 10 | Gemini 1.5 Flash (001) | 55.51 | #587 |
| 11 | Qwen2-Audio-7B-Instruct | 20.87 | #881 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-audio-mustard · How It Works · Data refreshed daily, snapshot 2026-10-24.