Fact · Languages and accents
99languages
The Whisper paper says its model predicts the spoken language with one token for each language in its training set, and that there are 99 such tokens, so the training data covers 99 languages.
Checked against its source on October 17, 2026
In the source's own words
“for each language in our training set (99 total)”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Section 2.3, Multitask format (language token list).
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- Whisper was trained on 680,000 hours of labeled audio. Each training segment starts with a language token, which the paper says is drawn from the VoxLingua107 model. The count describes the training set, not the accuracy of each language.
- What it does not show
- Having a language token means the model was trained on some audio in that language. It does not mean the model is accurate in all 99. The per-language error rates vary widely.