VoiceBench
Multimodal · 2024-10-22
VoiceBench evaluates LLM-based voice assistants on spoken instructions with real and synthetic speech variation. Scores use the official leaderboard Overall: the unweighted mean of nine subsets (AlpacaEval-full, CommonEval, WildVoice, SD-QA, MMSU, OpenBookQA, BBH, IFEval, AdvBench), with the three 1-5 judge scores multiplied by 20.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-Audio-2B | 73.1 |
| Phi-4-multimodal-instruct | 65.0 |
| LFM3-Omni-2B | 64.7 |
| LFM2.5-Audio-1.5B | 54.9 |