Fact · Languages and accents
The Whisper paper's authors found that most of the training data labeled as Welsh-to-English translation was actually English audio.
Checked against its source on October 17, 2026
In the source's own words
“Inspection shows the majority of supposedly Welsh translation data is actually English audio”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Discussion of the Fleurs translation results, after Figure 4.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The authors used this to explain why Welsh translation scored far lower than its supposed amount of training data predicted.
- What it does not show
- This is an inspection by the authors, reported as one example. It shows that the labels on translation training data can be wrong, which makes per-language scores harder to trust.