Is a lower CER always better than a lower WER?

They are not comparable. Quote one or the other consistently and state which you used, because the same transcript produces very different numbers under each.

Which should I use for English?

WER is the convention for English and most space-delimited languages. CER is the fallback when segmentation is ambiguous or when character-level accuracy is the thing being judged.

Does punctuation count toward CER?

It depends on the evaluation setup. Many benchmarks normalise punctuation and case away first, which is why published figures need their methodology stated to be meaningful.

Does CER work for English?

It can be calculated, but WER is the convention for English because word boundaries are unambiguous. CER is mainly reached for when segmentation is uncertain.

Definition

What Is Character Error Rate (CER)?

VideoText editorial team. Updated September 13, 2026.

Direct definition

Character error rate (CER) is the same edit-distance idea as WER, computed on characters instead of words. It is common for languages without clear word boundaries and as a complement to WER.

The sections below explain how character error rate is used in transcription and subtitle work, what people often mix up, and which VideoText workflow applies when the next step is a file or review pass.

Key takeaways

CER = (S + D + I) / number of reference characters.
A low WER can hide character-level mess in punctuation or morphology.
Do not convert CER to “percent accurate” without stating the metric.

When CER is preferred

Mandarin and other scripts are often scored with CER. Some punctuation studies also report CER because a period is a character, not a word.

How CER is calculated

CER counts the character insertions, deletions and substitutions needed to turn the recognised text into the reference, divided by the number of characters in the reference. It is the same edit-distance idea as word error rate, measured one level down.

Because the denominator is much larger, CER values are numerically smaller than WER on the same transcript. A 10% WER and a 3% CER can describe the same file, so the two are never compared directly.

Where CER is the better measure

Languages written without spaces — Japanese, Chinese, Thai — have no unambiguous word boundary, so word-level scoring depends on how the text was segmented and becomes an artefact of the tokenizer rather than a property of the transcript.

CER also suits work where partial correctness matters: a name misspelled by one letter scores as a total loss under WER but as a small error under CER, which is closer to how a reader experiences it.

Frequently asked questions

Is a lower CER always better than a lower WER?: They are not comparable. Quote one or the other consistently and state which you used, because the same transcript produces very different numbers under each.
Which should I use for English?: WER is the convention for English and most space-delimited languages. CER is the fallback when segmentation is ambiguous or when character-level accuracy is the thing being judged.
Does punctuation count toward CER?: It depends on the evaluation setup. Many benchmarks normalise punctuation and case away first, which is why published figures need their methodology stated to be meaningful.
Does CER work for English?: It can be calculated, but WER is the convention for English because word boundaries are unambiguous. CER is mainly reached for when segmentation is uncertain.

Related terms

Sources

OpenAI: Whisper paper (evaluation / normalizer context)