Fact · Languages and accents
0.83squared correlation
Across languages, the log of Whisper's word error rate has a squared correlation of 0.83 with the log of the amount of training data for that language.
Checked against its source on October 16, 2026
In the source's own words
“We find a strong squared correlation coefficient of 0.83 between the log of the word error rate and the log of the amount of training data per language.”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Section 3.3.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The correlation is across the Fleurs languages and Whisper's training data per language.
- What it does not show
- Correlation, not proof of cause. The paper also notes outliers, such as languages whose scripts or linguistics differ from the bulk of the training data.