Word error rate (WER)

Word error rate (WER) is the most common way to compare speech-to-text models. The model's output is compared with a correct reference text written by a person who listened to the recording, and the differences are counted. The lower the rate, the more accurate the transcript. It is handy because it is a single number, but it says nothing about which words were misheard or how much those mistakes matter.

How it works

The formula is (substituted + deleted + inserted words) ÷ number of words in the reference. Before scoring, both texts are normalized: case, punctuation and the way numbers are written ("15" versus "fifteen") are made consistent. Because these normalization rules change the result considerably, figures published by different sources can't be compared directly. In agglutinative languages such as Turkish, a single wrong suffix makes the whole word count as an error, so character error rate (CER), which works at the letter level, is often reported alongside it.

Why it matters

Published figures are mostly measured on clean, high-quality recordings. Phone lines, background noise, accents and domain-specific terms (drug names, product codes, street names) push the error rate up in real calls. When choosing a voice agent or a call transcription product, test with recordings that resemble your own calls and look at the details that matter rather than the overall rate: were names, dates, phone numbers and amounts understood correctly? For a practical test plan, see our checklist for testing a Turkish AI phone assistant.