AV-SpeakerBench
Multimodal · 2025-12-01
AV-SpeakerBench is a curated benchmark of 3,212 four-choice questions on speaker-centric audiovisual reasoning in about 2,051 real-world videos, where answers depend on who speaks, what is said and when. Models receive video with its soundtrack, and scores are multiple-choice accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-Omni-2B | 46.8 |
| Phi-4-multimodal-instruct | 38.0 |