Character error rate is word error rate computed on characters instead of words. Autrans' language page figures are word-based.
Character error rate is word error rate computed on characters instead of words: substitutions, deletions and insertions of characters divided by the number of characters in the reference. It is gentler than WER on a near-miss spelling, and it shows whether errors are small slips or wrong words.
Also searched as "cer".
Autrans' language page figures are word-based. CER matters for languages whose spelling depends on marks and special letters (tone marks in Yoruba and Igbo, special letters in Hausa and Twi), where a missing mark is a single-character slip but a whole wrong word under WER. Autrans' own app evaluation scripts test diacritic restoration on Yoruba and Igbo outputs.
Reference: 'ọkọ' (3 characters, with the dot-below letter ọ stored as one character, Unicode U+1ECD). Transcript: 'oko'. Two of three characters are substituted, so the character score is 2 divided by 3, about 67 in 100. The same single word is wrong, so the word score is 1 divided by 1, a whole word. If the text is stored with a combining dot instead (NFD), the reference is 5 code points and the same slip counts as 2 deletions, the character score is 2 divided by 5, 40 in 100: normalise text before scoring.
Character error rate is word error rate computed on characters instead of words: substitutions, deletions and insertions of characters divided by the number of characters in the reference. It is gentler than WER on a near-miss spelling, and it shows whether errors are small slips or wrong words.
Autrans' language page figures are word-based. CER matters for languages whose spelling depends on marks and special letters (tone marks in Yoruba and Igbo, special letters in Hausa and Twi), where a missing mark is a single-character slip but a whole wrong word under WER.
Both scores count substitutions, deletions and insertions against a reference and divide by its length. The full answer, with sources, is on the page for WER vs CER.
It rests on jitsi.github.io, jiwer and lightning.ai, char error rate. What each source supports is listed above, next to its link.
Reference: 'ọkọ' (3 characters, with the dot-below letter ọ stored as one character, Unicode U+1ECD). Transcript: 'oko'. Two of three characters are substituted, so the character score is 2 divided by 3, about 67 in 100. The same single word is wrong, so the word score is 1 divided by 1, a whole word. If the text is stored with a combining dot instead (NFD), the reference is 5 code points and the same slip counts as 2 deletions, the character score is 2 divided by 5, 40 in 100: normalise text before scoring.
Send Autrans into your next meeting, use it in the room, or upload a recording. Your first 30 minutes are free, given once, and no credit card is needed to start.
Last checked 22 October 2026. How we measure accuracy. Spotted a mistake? Write to hello@autrans.online.