Glossary · Speech
What is character error rate (CER)?
Character error rate (CER)
Character error rate, or CER, scores a speech recogniser letter by letter instead of word by word. A small spelling slip costs a little, where word error rate would count the whole word as wrong.
On a call
The caller said "पहुँचेगा" and the recogniser wrote "पहुंचेगा", with a dot where the chandrabindu should be. WER counts a whole wrong word. CER counts one character out of eight.
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
CER works exactly like word error rate, one level down. Line up the reference and the recogniser’s output character by character, count substitutions, deletions and insertions, and divide by the length of the reference.
It’s the standard score for languages where “word” is a slippery idea. Mandarin and Japanese are written without spaces, so researchers score them in characters. It has become just as useful for Hindi, for a different reason: spelling.
Devanagari gives you several correct-looking ways to write the same sound. A nasal can be a bindu or a chandrabindu. A borrowed English word can be spelled more than one way, and a nukta under a letter comes and goes. Transcribers disagree with each other on these, and WER punishes every disagreement as a full word wrong.
CER is gentler. A one-letter spelling difference in a long word costs a fraction of a word. So when WER looks high but CER is low, the recogniser is often hearing the right words and spelling some of them differently. When both are high, it’s more likely mishearing.
That makes the pair more useful than either number alone. Read them together, and look at a sample of the errors before you trust the headline.
CER has blind spots too. Swap one digit in a phone number and CER barely moves, yet the call has failed. For voice agents, the words that carry the job still need checking on their own.
Like WER, CER depends on normalisation. Decide how to treat punctuation, digits and Roman-script English before you score, and apply the same rules to every system you compare.
In Voice Infra
This is what Voice Infra does about it.
We report CER next to WER for Mira STT, so a reader can tell spelling noise from real misses. On the RinggAI Hindi benchmark its CER was 3.21%, with the full method and the test sets where we trail on the Mira STT page.