Glossary
Character error rate (CER)
Character error rate (CER) is the number of character insertions, substitutions and deletions needed to turn a recognizer's output into the correct text, divided by the number of characters in the correct text. It is the usual measure where words are not separated by spaces.
Updated October 17, 20264 cited facts, each traced to its source
Also called CER.
Character error rate (CER) measures how often a transcript gets individual characters wrong. It counts the minimum number of insertions, substitutions and deletions needed to turn the recognizer's output into the reference text, then divides by the number of characters in the reference. A lower number is better, and zero means the output matches the reference exactly.
How the count works
The method is a Levenshtein distance computed at the character level. A 2024 Mandarin challenge describes CER as counting the minimum number of insertions, substitutions and deletions required to transform the recognizer output into the reference 1. An OCR study states the same idea as the minimum number of character substitutions, deletions and insertions needed to turn the ground truth into the output, divided by the number of characters in the ground truth 2.
A small example shows the logic. If the reference is "ten" and the output is "tent", the output has one inserted character, so the count is one error. The rate then depends on how many characters the whole reference contains.
Why CER is used for Chinese and Japanese
Word error rate needs word boundaries, and Chinese and Japanese do not put spaces between words. The Whisper paper put a space between every letter for Chinese, Japanese, Thai, Lao and Burmese, which means it measured character error rate for those languages instead of word error rate 3. For Mandarin, the AISHELL-1 baseline reported a character error rate of 7.62 percent on the test set, from the best of three 2017 models 4. Those numbers are read speech in a quiet setting, so they show what is possible, not what you will get at a noisy desk.
What a CER number does not tell you
A CER value is only comparable with other values on the same test set, with the same normalization of punctuation and numbers. A lower CER on one corpus does not promise the same result for your voice or your topic. CER also treats every character as equal, so one wrong character in a name counts the same as one wrong common word.
Why it matters for dictation and typing
If you dictate Chinese, Japanese or Korean, look for CER results on a test set similar to your speech. For those languages, CER is closer to how a reader judges a transcript than a word count would be. Many dictation products do not publish CER, so compare tools by testing your own speech. The Japanese, Chinese and Korean speech to text guide covers each language and the typing side.
Related terms
- Word error rate: the same method, counting words instead of characters.
- Input method editor (IME): how characters are typed in these languages.
- Low-resource language: languages where CER is often much higher.
Guides that cover it
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
Related terms
- Dictation Languages, Accents and Voice TranslationThe topic this term belongs to
- AccentAn accent is the way a speaker pronounces a language, shaped by where they grew up, their first language or their community. Accents change pronunciation, not vocabulary, and speech recognition error rates can differ sharply from one accent to another.
- Code-switchingCode-switching is the habit of moving between two or more languages within a conversation or a single sentence. Speech recognizers built around one language often misread the switched words, so mixed speech is harder to dictate than speech in one language.
- DiacriticsDiacritics are marks added to a letter, such as the acute accent in é, the tilde in ñ or the vowel marks in Arabic, that change how a letter sounds or which word it spells. Speech tools and everyday writers often leave them out, so a dictated word can come out spelled differently from what you meant.
- DialectA dialect is a regional or social variety of a language, with its own vocabulary, grammar and pronunciation. Speech recognition is usually less accurate on dialects than on the standard form, because labeled training data for dialects is often scarcer.
- Input method editor (IME)An input method editor (IME) is software that supports text input for characters that cannot be typed easily on a standard keyboard, such as many East Asian characters. Speech input can also be an input method, and Android allows only one input method to be enabled at a time.
- LocaleA locale is a language combined with a region, and sometimes a script, such as en-GB for English as used in the United Kingdom. Dictation and speech settings use locales because the same language can differ in spelling, punctuation and speech patterns from one region to another.
- Low-resource languageA low-resource language is one with little digital speech and text data for training speech recognition and other language tools. There is no fixed cutoff, and a language can be well resourced for one task and poorly resourced for another.
- Modern Standard Arabic (MSA)Modern Standard Arabic (MSA), also called Fusha, is the standardized form of Arabic used in writing and formal speech, such as news broadcasts and official documents. Speakers use regional dialects in everyday conversation, and speech tools recognize formal Arabic more accurately than dialect speech.
- Non-native speechNon-native speech is speech in a language that is not the speaker's first language, often called L2 speech or foreign-accented speech. Recognizers can transcribe read non-native English well, but they handle hesitations, repetitions and revisions differently, and results depend on the system.
- PinyinPinyin, also called Hanyu Pinyin, is the system that spells Mandarin Chinese syllables with the Latin alphabet and marks the tones with diacritics. People type Chinese by spelling the sound in Pinyin and choosing the character they want, and speech tools must choose among characters that share one Pinyin spelling.
- Right-to-left (RTL) scriptA right-to-left (RTL) script is a writing system, such as Arabic, Hebrew, Persian or Urdu, whose letters run from right to left. Text that mixes these letters with digits or Latin words is bidirectional, so software needs a specific algorithm to order and display it.
- Speech translationSpeech translation converts speech in one language into the same meaning in another language, as text or as synthesized speech. Most systems first recognize the speech and then translate the text, so a recognition error carries into the translation.
Pages that use this term
- Dictation Languages, Accents and Voice TranslationTopic
- Japanese, Chinese and Korean Speech to TextGuide
- How to Translate Audio and Video to English TextGuide
- Transcribe and Translate Audio in Other LanguagesGuide
- How can you type Japanese faster?Question
- How do you dictate in Japanese on a phone or computer?Question
- How do you transcribe Japanese, Chinese or Korean audio?Question
- How do you translate Japanese audio to English?Question
- How do you voice type in Korean?Question
- How fast do people type in Japanese?Question
- What is the best speech to text for Japanese?Question
- Input method editor (IME)Glossary
- Low-resource languageGlossary
- PinyinGlossary