Word error rate calculator
Paste the correct text and a machine transcript. You get the rate every vendor quotes — and, underneath it, where the errors actually landed. It runs in your browser.
The correct transcript
The transcript to score
Why the breakdown matters
Word error rate is the count of substituted, deleted and inserted words divided by the length of the correct transcript. It is a real measurement, and it treats every word as equally important.
That is the problem. Consider two transcripts of the same hour. One scores 3%, and every error is a filler word or a false start. The other scores 1%, and its errors are a surname, a figure in a budget discussion, and a dropped “don’t”. The second wins the benchmark by a factor of three. The first is the one you can publish; the second quotes someone as agreeing to something they refused.
So the number above is the one you would see on a marketing page, and the table under it is the one that decides whether the transcript is any use. We wrote the longer version of this argument up in What “99% accurate” actually means.
How it counts
- Case and punctuation are ignored
- As the metric is normally defined. “The cat sat.” and “the cat sat” score as identical.
- Names are capitals found mid-sentence
- A cheap approximation, and an honest one: every sentence starts with a capital, so a sentence-initial one tells you nothing and isn't counted.
- Numbers count digits and number words
- “240”, “eleven” and “percent” all land in the same bucket.
- Each word goes in exactly one bucket
- Most consequential first: a dropped “no” is counted as a negation whether or not it was also capitalised.
- Insertions have no bucket
- The categories are counted over the correct transcript, and an invented word has no counterpart there. It still counts towards the rate.
We don't publish an accuracy number
We could pick a favourable benchmark and generate a good one, the way everybody does. It would tell you nothing about your recording. What txscribe does instead is make its errors findable: every word carries a timestamp, so you click a line and hear whether it is right.
See what it does