Fact · Speech recognition
5.2percent word error rate
On the LibriSpeech test-other split, the Whisper large-v2 model had a word error rate of 5.2 percent, in the paper's English table with greedy decoding.
Quote found word for word on the live source, last checked October 19, 2026
In the source's own words
“Whisper large-v2 2.7 5.2 4.0 3.9 17.6 13.8 9.0 6.2 16.2 25.5 16.9 36.4 7.3 4.4”
Robust Speech Recognition via Large-Scale Weak Supervision (full text), Alec Radford et al., OpenAI (arXiv), 2022, Table D.1.1 (English transcription, greedy decoding), Whisper large-v2 row, column LibriSpeech test-other.
Open the original (DOI 10.48550/arXiv.2212.04356)- What it measures
- LibriSpeech test-other is one of the two LibriSpeech test splits the paper evaluates. The English table lists the same model on many other test sets.
- What it does not show
- Greedy decoding, not beam search. Read audiobook speech, so it is a clean read-speech condition, not conversation.
In the data
A row of Speech Recognition Accuracy: Word Error Rates, as Whisper large-v2, LibriSpeech other: 5.2 percent.
- System
- Whisper large-v2
- Kind
- Open model
- Test set
- LibriSpeech test-other
- Language
- English
- Speech
- Clean read speech
- Year
- 2022