Character error rate
In one sentence
Character error rate (CER) is transcription accuracy measured over individual characters rather than whole words, using the same substitution, deletion and insertion formula as word error rate, and it is the right metric where word boundaries are unclear or where exact spelling matters.
Definition
Character error rate measures transcription accuracy at the level of individual characters rather than whole words.
It uses the same formula as word error rate, counting substitutions, deletions and insertions divided by the reference length, applied to characters instead of words.
Character error rate is used where word boundaries are unclear and where exact spelling matters. It is diagnostically useful even on languages where word error rate is the headline metric, because it separates two failures that word error rate blurs together.
Where character error rate is the appropriate metric
- Languages without whitespace word segmentation, such as Chinese, Japanese and Thai, where word error rate requires an arbitrary tokenization decision that changes the result.
- Content where exact spelling matters: reference codes, license plates, part numbers, email addresses.
- Evaluating near-miss quality. A word transcribed with one character wrong scores as a complete failure under word error rate and as a near-success under character error rate.
What character error rate reveals that word error rate hides
- A system producing phonetically close but misspelled proper nouns has poor word error rate and good character error rate. That tells you the acoustic modeling is working and the lexicon is missing the term.
- That distinction points straight at custom vocabulary as the fix, rather than at a model change.
- So character error rate earns its place as a diagnostic even where word error rate is the number on the spec sheet.
Where character error rate misleads
- It is insensitive to word-level meaning. Transcribing cat as cot is one character and a complete change of meaning.
- For most English conversational applications, word error rate correlates better with task success.
- Character error rate on long words looks better than on short ones for the same practical error rate, which makes cross-corpus comparison unreliable.
The practical recommendation
- Report word error rate as the headline for word-segmented languages.
- Report character error rate alongside it when diagnosing whether errors are lexicon problems or acoustic ones.
- Use character error rate as the primary metric for non-segmented languages and for identifier-heavy content.
Common misconception
That character error rate is simply a finer-grained word error rate and therefore strictly better. It measures a different thing and is not universally more informative.
Why it matters commercially
Character error rate is the right metric for identifier-heavy content, which is common in business voice applications, and it is a useful diagnostic for telling a lexicon gap apart from an acoustic failure. Getting the diagnosis right decides whether you fix the vocabulary or replace the model.
In voice specifically
In text the visitor types an exact reference code and it is captured exactly. In speech the code has to survive recognition, and a single wrong character can break a lookup that a whole correct word would have passed. Character error rate is the measure that sees that near-miss for what it is, where word error rate would call the whole word a failure and hide which kind of failure it was.
Where AsqVox fits
When an AsqVox agent handles reference numbers or unusual product codes pulled from uploaded documents, character-level accuracy is the measure that matters, and the text chat fallback is often the better way to capture an exact identifier than the spoken channel.
Visual
The same transcript, two verdicts
Not a finer word error rate. A different question.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Character error rate applies the same substitution, deletion and insertion formula as word error rate, computed over characters instead of words, and it is the standard accuracy metric for languages without whitespace word segmentation.
Same formula, over charactersindustry rangeStandard speech-evaluation definition, 2026 - A definition rather than a benchmark. It is the right primary metric for Chinese, Japanese and Thai, where word error rate needs an arbitrary tokenization decision that changes the result.
Whisper Large-v3 records around 2.7 percent word error rate on clean LibriSpeech test-clean.
2.7% WER, cleanindependentWhisper Large-v3 on LibriSpeech test-clean, 2026 - A word-level figure. Character-level error on the same audio is typically lower and is not directly comparable to it.
The same model lands at roughly 8 to 12 percent word error rate on real-world English audio.
8 to 12% WER, real worldindustry rangeReported real-world word error rate for Whisper Large-v3, 2026 - Reported rather than measured once, and still a word-level figure. The character-level number on the same audio would sit lower.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate in production conditions.
5.26% to 6.84% WERvendor claimDeepgram Nova-3 reported figures, 2026 - Published by the provider, and a word-level figure. Providers rarely publish the character-level number beside it for English.
There is no standard convention for text normalization in reported character error rate, the same gap that undermines word error rate reporting, which limits cross-provider comparison.
-no reliable figureTwo providers can both report a character error rate honestly and still not be comparable, because they normalized the text differently before scoring.
Providers rarely publish character error rate alongside word error rate for English, which removes a useful diagnostic signal from public benchmark material.
-no reliable figureThe signal being removed is the one that separates a lexicon gap from an acoustic failure, which is exactly the diagnosis a buyer with identifier-heavy content needs.
Examples
In practice
A team sees poor word error rate on a corpus full of product codes and concludes the model is unsuitable. Running character error rate beside it shows very low character error, which says the acoustics are being captured accurately and the codes are being spelled almost right. The fix is custom vocabulary for the code format, not a different provider, and the whole diagnosis came from the second metric.
The everyday version
Character error rate counts mistakes letter by letter instead of word by word. If the system hears your business name and gets a single letter wrong, word-level scoring marks the whole word a failure while letter-level scoring shows it came out nearly right. That gap tells you the problem is a missing dictionary entry, not a bad system.
Usage
Who says it
- Speech researchers and engineers, especially those working with languages that do not separate words with spaces.
- It appears in model documentation for multilingual systems.
Where it turns up
- In a spec sheet, next to accuracy metrics, the list of supported languages and the evaluation methodology.
Common misuse
- Comparing character error rate across corpora with different average word lengths, where the same practical error rate scores better on long words than on short ones.
- Treating it as universally more informative than word error rate. It measures a different thing and is not always the better one.
- Reporting it for English conversational content, where word error rate tracks task success more closely.
Questions people ask
What is the difference between character error rate and word error rate?
Both use the same formula, counting substitutions, deletions and insertions divided by the reference length, but character error rate applies it to individual characters and word error rate applies it to whole words. A word transcribed with one letter wrong is a complete failure under word error rate and a near-success under character error rate. Word error rate is the usual headline for languages that separate words with spaces; character error rate is primary for languages that do not, and for identifier-heavy content.
When should you use character error rate?
Use it as the primary metric for languages without whitespace word segmentation, such as Chinese, Japanese and Thai, where word error rate needs an arbitrary tokenization choice that changes the result. Use it for content where exact spelling matters, like reference codes, license plates, part numbers and email addresses. And report it alongside word error rate when you are diagnosing whether errors come from a missing lexicon or from the acoustic model.
What does good character error rate with poor word error rate mean?
It means the acoustic modeling is working and the lexicon is missing the term. The system is capturing the sound accurately and spelling an unfamiliar word almost right, so it loses whole words on word error rate while barely moving character error rate. The signal points straight at custom vocabulary as the fix, rather than at a different model or provider.
Is character error rate just a finer word error rate?
No. It measures a different thing and is not universally more informative. It is insensitive to word-level meaning, so cat transcribed as cot is one character and a completely different animal, and for most English conversational content word error rate tracks task success more closely. Character error rate on long words also looks better than on short ones for the same practical error rate, which makes cross-corpus comparison unreliable.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.