Character error rate

Character error rate, usually shortened to CER, measures how well software turns a picture of text into actual text. It counts the characters that were wrong, missing or invented and divides by the number of characters that should have been there. Lower is better, so a CER of 2.4 percent means roughly 24 characters in every thousand came out wrong, or 97.6 percent accuracy.

It is the number to compare when choosing a text recognition tool, and it is stricter than its cousin word error rate, because a single wrong letter counts rather than writing off the whole word. Two warnings apply. CER is always measured on a particular set of pages, so results on clean modern print say little about faded scans or handwriting. And it treats a harmless modernisation, such as turning an old long s into a normal s, as an error, which is why the same model can look bad for scholarship and fine for AI training.