Word error rate
The standard accuracy measure for speech-to-text: substitutions, deletions and insertions divided by the number of words in a correct reference transcript.
Why it matters. Lets you compare transcription accuracy, as long as it is measured on audio like yours.
How it works
Word error rate (WER) compares a transcript with a correct reference: (substitutions + deletions + insertions) divided by the number of words in the reference. Lower is better, and 0% is a perfect match; because insertions count, WER can exceed 100%.
Results depend on the test audio, the language and how text is normalized before comparing (casing, punctuation, numbers as digits or words). Makers report WER on different datasets, so numbers are comparable only on the same test set with the same normalization.
Measure WER on a sample of your own audio with carefully checked reference transcripts. Also look at which words are wrong: a mistake in a drug name or account number matters more than a missed 'um'.
Worked example
Illustrative: against a 200-word reference transcript, a model substitutes 6 words, drops 3 and adds 1 extra. WER = (6 + 3 + 1) / 200 = 5%.
Common mistakes
- Comparing WER figures measured on different datasets or with different normalization.
- Treating all errors as equal when names, numbers and domain terms matter most.
- Measuring on clean read speech when your audio is noisy, accented or conversational.
See it on a real card: Benchmarks on Veo 3.1.