Fact · Languages and accents
16xincrease in training data per halving
The Whisper authors estimate that a language's word error rate halves for every 16-fold increase in its training data.
Checked against its source on October 16, 2026
In the source's own words
“results in an estimate that WER halves for every 16”
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 estimate comes from a log-log regression of word error rate on hours of recognition training data per language, across the Fleurs languages.
- What it does not show
- A trend line with outliers, fitted to one model family. It describes Whisper, not every recognizer.
Source
Cited on
- Dictation Languages, Accents and Voice TranslationTopic
- Speech Recognition Models: How Neural ASR WorksGuide
- Whisper Supported Languages and Accuracy by LanguageGuide
- Transcribe and Translate Audio in Other LanguagesGuide
- How good is Whisper at Hindi?Question
- How good is Whisper at Korean?Question
- Why does Whisper transcribe in the wrong language?Question