What is the difference between WER and character error rate?

WER and CER use the same arithmetic, but CER counts characters instead of words

Word error rate and character error rate are the same calculation applied to two units. Both align the output with a reference text, count the substitutions, deletions and insertions needed to match them, and divide by the length of the reference. WER uses words as the unit, and CER uses characters. The guide to word error rate explains the formula in full, so this page focuses on the choice between the two.

The same definition appears in different fields. In a study of OCR on Sinhala and Tamil text, character error rate is based on Levenshtein distance and counts "the minimum number of character-level operations (substitutions, deletions, and insertions)" needed to turn the ground truth into the OCR output ¹. In the 2024 Mandarin stuttering challenge, CER was defined as the minimum number of insertions, substitutions and deletions needed to turn the recognizer output into the reference ². Both are the same edit distance, applied to characters.

Use CER when the language is written without spaces between words

WER needs a word boundary for every token, and Mandarin text has none. A scorer must first split the characters into words, and that segmentation is itself uncertain. A 2013 study by Liu and colleagues reports that native speakers agree on where word boundaries fall only about 75 percent of the time ³. The authors cite that level of agreement as the reason character error rate is the commonly used metric for Mandarin speech recognition.

The same reasoning applies to other languages written without spaces. If you score them with WER, the word count depends on a segmenter, and two segmenters can give two different scores for the same output. CER avoids that choice, because a character is the same unit in the reference and in the output.

One wrong letter can cost a whole word, so WER is usually higher than CER

In WER, a word is either right or wrong, so one misspelled letter makes the whole word an error. CER counts that letter as one edit out of all the characters in the text. The effect is visible in a 2025 study of OCR on typewritten historical records. Its OCR output had a character error rate of 1.08 percent and a word error rate of 5.06 percent ^{4 5}. The same text is scored with both units, and the word figure is several times larger.

The gap is not a fixed rule. It depends on how often errors fall inside words and how long the words are. The study is one collection of documents, so use it to see the effect, not to predict the ratio for your audio.

It depends on the language, the task and the reader

For English speech, WER is the default, and it matches how readers judge a transcript: a wrong word is a visible mistake. For Mandarin, CER is the convention, and the word-level score is only meaningful after you state the segmenter. For OCR, report both, because a reader who corrects a document cares about the characters and the words they must fix.

Never compare a CER with a WER. A CER of 5 percent and a WER of 5 percent measure different units over different denominators, so they do not describe the same quality. A comparison is fair only when both systems are scored with the same unit, on the same audio, after the same normalization.

If your task needs word-level results in an unspaced language, for example keyword search or word timing, you will need a word-based score as well. Then report the segmenter and its version next to the number, because the segmenter changes the result.

What to do

1. Pick the unit that matches the language: WER for text with spaces between words, CER for text without them.
2. Write down the normalizer for each unit, covering case, punctuation and character set. The guide to normalization explains why this changes scores.
3. Name the unit in every report, for example "CER (characters)" or "WER (words)", and never leave it implied.
4. For OCR, report both CER and WER, with the number of reference characters and words.
5. For Mandarin, report CER first. Add a word-level score only when the task needs one, and state how the words were segmented.

Related reading: what counts as a good word error rate, how to calculate word error rate and whether word error rate is the same as accuracy. For the wider picture, see the overview of speech recognition.