1,205
±119 95% CI
64%
Win rate
72
Battles
AudioLab Leaderboards
Pairwise preference rankings and independent benchmarks for leading speech models.
Valid votes
190
Models
6
Method
Bradley–Terry MLE
Speech to Speech rankings
Models are ordered by rating. Models are marked preliminary until 200 valid battles; disabled models are hidden. Newly added models appear unranked at the bottom until they record their first battles.
Controls only for which response was heard first.
Preliminary models have fewer than the required valid battles or a weakly connected matchup graph. Treat their rank and confidence interval as directional.
1,183
±129 95% CI
61%
Win rate
52
Battles
1,123
±119 95% CI
52%
Win rate
65
Battles
1,115
±113 95% CI
50%
Win rate
61
Battles
1,037
±99 95% CI
39%
Win rate
66
Battles
1,000
±0 95% CI
33%
Win rate
63
Battles
Performance by language
Win rate per language. Cells with too few battles are left blank.
| Model | English |
|---|---|
| Google Gemini 2.5 Flash Native Audio | 66% |
| OpenAI GPT Realtime | 54% |
| OpenAI GPT-4o Audio | 50% |
| OpenAI GPT Realtime 1.5 | 57% |
| xAI Grok Voice | 38% |
| Alibaba Qwen3-Omni Flash | 31% |