Fact · Languages and accents
125,000hours
Whisper's training data includes 125,000 hours of audio for translating speech from other languages into English.
Checked against its source on October 17, 2026
In the source's own words
“The dataset also includes 125,000 hours of X→en translation data”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Introduction.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- The paper's training set mixes speech recognition and translation data, and the translation part is what lets Whisper's translate task work without a second model.
- What it does not show
- The paper describes this data as noisy, and some of the labels are wrong, as the Welsh example in the same paper shows.