HELM Audio - Corebench: leaderboard
The "Corebench" 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: PEM. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 05-06) | 81.3 | #148 |
| 2 | Gemini 1.5 Pro (002) | 79.9 | #350 |
| 3 | Gemini 1.5 Flash (002) | 77.6 | #585 |
| 4 | Gemini 2.0 Flash | 75.6 | #344 |
| 5 | Gemini 2.0 Flash (Preview) | 75.4 | #366 |
| 6 | Gemini 2.0 Flash Lite | 73.7 | #453 |
| 7 | Gemini 1.5 Flash (001) | 72.2 | #580 |
| 8 | Gemini 1.5 Pro (001) | 65.9 | #446 |
| 9 | Gemini 2.5 Flash (Preview 05-20) | 64.4 | #214 |
| 10 | Qwen2.5-Omni-7B | 56 | #632 |
| 11 | Qwen2-Audio-7B-Instruct | 23.3 | #888 |
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-corebench · How It Works · Data refreshed daily, snapshot 2026-10-23.