VoxPopuli-Cleaned-AA
Classic NLP · 2026-02-12
VoxPopuli-Cleaned-AA is Artificial Analysis's manually corrected 628-record subset of the English VoxPopuli test data for speech-to-text evaluation. Scores are audio-duration-weighted word error rate after Whisper-style text normalization; lower is better.
Top models (lower is better)
| Model | Score |
|---|---|
| Interfaze Beta | 2.4 |
| Gemini 3 Flash Preview | 4.0 |
| Gemini 3.5 Flash | 4.0 |