HalluAudio - Speech: leaderboard
Hallucination detection in the speech domain of HalluAudio (2026): human-verified audio question-answer pairs with adversarial prompts, mixed audio and unanswerable questions probe whether an audio-language model reports sounds or content that are not there. Unweighted mean accuracy over ten task types, in percent; higher is better.
Metric: Average classification accuracy (%) over the ten speech task types of HalluAudio (human-verified audio question-answer pairs with adversarial, mixed-audio and unanswerable prompts; zero-shot; outputs normalized to labels, refusals counted wrong); unweighted mean of the task accuracies; higher is better. Source: arxiv.org. Saturation forecast: Around 2032. 11 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiMo-Audio-7B-Instruct | 52.84 | #1336 |
| 2 | GPT-4o Audio (Preview) | 45.51 | #793 |
| 3 | LLaMA-Omni2-7B | 40.94 | |
| 4 | Kimi-Audio-7B | 40.73 | |
| 5 | Step-Audio-2-mini | 38.29 | |
| 6 | Phi-4 Multimodal Instruct | 35.47 | #4154 |
| 7 | Qwen-Audio-Chat | 32.16 | |
| 8 | Qwen2-Audio-7B | 28.72 | |
| 9 | Llama 3.1 8B Omni | 18.26 | |
| 10 | Pengi | 16.72 | |
| 11 | Gemini 2.5 Flash | 16.6 | #463 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=halluaudio-speech · How It Works · Data refreshed daily, snapshot 2026-11-21.