Word error rate calculator

Check how accurate a transcript is. Paste the words that were actually spoken and the transcript you want to score. Your text stays in your browser: nothing is uploaded.

How word error rate is calculated

WER = (S + D + I) / N. S is words swapped for another word, D is words missed, I is words added that were not said, and N is the number of words in the reference. The calculator finds the smallest set of edits that turns the reference into the transcript (a word-level edit distance) and counts each kind. 5% means about 1 word in 20 needs a fix. WER can go above 100% when a transcript adds many words.

• Cleanup matters. With the box ticked, both texts are lowercased and punctuation and symbols are removed first, the same cleanup our published benchmark uses for non-English languages. Untick it to count "Hello," and "hello" as different words.
• Numbers and spellings. "10" and "ten", or "color" and "colour", count as different words here. Research benchmarks often unify them first, which lowers the score. Compare numbers only when both used the same cleanup.
• Limit. Up to 5,000 words per text. Score long recordings in parts and add up S, D, I and N before dividing.

What a good WER looks like

On clean read speech, good engines score under 5%. Real meetings, phone calls and noisy rooms score higher for every engine. Our measured results on public test sets, with every transcript we scored, are on the speed and accuracy benchmark page.

Link to this calculator

Teaching speech recognition or writing a tool review? Use the calculator freely and paste this link into your page.

<a href="https://convertaudiototext.com/tools/wer-calculator">Word error rate calculator</a> by ConvertAudioToText: score any transcript against a reference, free in the browser.