SpeakerSleuth - Discrimination: leaderboard
Speaker discrimination task of SpeakerSleuth (Lee, Pyo, Seo and Jo, Seoul National University, ACL 2026): with one of a speaker's five dialogue turns masked, an audio-language model picks which of three shuffled candidate clips (the original, a voice-converted similar speaker, a voice-converted opposite-gender speaker) belongs to the speaker. Accuracy in percent over 606 items, where chance is 33.3; higher is better.
Metric: Classification accuracy (%; 606 three-candidate items). Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 81.5 | #262 |
| 2 | Gemini 2.5 Flash | 75.6 | #452 |
| 3 | Gemini 2.5 Flash Lite | 46.5 | #958 |
| 4 | Qwen2.5-Omni-7B | 42.7 | #1797 |
| 5 | Phi-4 Multimodal Instruct | 39.9 | #3914 |
| 6 | Gemma 3n E4B | 37.1 | #3330 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=speakersleuth-discrimination · How It Works · Data refreshed daily, snapshot 2026-11-12.