Omni-Persona - Audio-to-Audio: leaderboard
Audio-to-Audio persona matching on Omni-Persona (2026): a query voice clip is matched to personal audio to answer personalized questions, scored by calibrated accuracy, which combines judged answers on answerable items with correct abstention on items whose persona is absent from the context. Percent; higher is better.
Metric: Calibrated accuracy (%): item-weighted accuracy over answerable items (answer judged correct by a GPT-5.4-mini judge) and absent-persona items (correct abstention by keyword matching), on the audio-to-audio persona matching group (query voice clip, personal audio context), zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Gemma 4 E2B (IT) | 31.2 | #2451 | |
| 2 | MiniCPM-o-4.5 | Thinking | 29.1 | #1252 |
| 3 | Qwen2.5-Omni-3B | 25.9 | #3519 | |
| 4 | Gemini 3.6 Flash | Medium | 23.9 | #81 |
| 5 | Interactive-Omni-8B | 21.5 | ||
| 6 | Phi-4 Multimodal Instruct | 20.6 | #4154 | |
| 7 | Qwen2.5-Omni-7B | 17 | #1884 | |
| 8 | gemma-4-E4B-it | 16.6 | #1539 | |
| 9 | Qwen3 Omni 30B A3B Instruct | 15.8 | #930 | |
| 10 | Gemini 3.5 Flash Lite | 10.1 | #299 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=omni-persona-audio-to-audio · How It Works · Data refreshed daily, snapshot 2026-11-21.