Fact · Speech recognition
33.0percent word error rate
On the FLEURS benchmark, the Whisper large-v2 model had a word error rate of 33.0 percent for Welsh.
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“Whisper large-v2 36.7 140.3 16.0 106.2 23.4 45.4 14.6 104.1 15.7 7.3 14.7 13.3 33.0 13.8”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Table 13 (FLEURS), Whisper large-v2 row, column Welsh.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- FLEURS is the paper's multilingual read-speech benchmark. Its table gives each language's word error rate in a column, and the Whisper large-v2 row is the largest of the table's models.
- What it does not show
- Read speech on one test set per language, so it describes clean read speech, not conversation or noise. Word error rate depends on the paper's text normalization. Only the large-v2 row is written here.
In the data
A row of Speech Recognition Accuracy: Word Error Rates, as Whisper large-v2, Welsh: 33 percent.
- System
- Whisper large-v2
- Kind
- Open model
- Test set
- FLEURS
- Language
- Welsh
- Speech
- Clean read speech
- Year
- 2022