Word error rate (WER)

Word error rate (WER) is the standard measure of how accurately a speech recognizer transcribes speech. It is the proportion of word errors to the words in the correct transcript. Researchers have used the measure since the early days of the field: a 2004 review describes it as the de facto standard index for speech recognition assessment ¹.

WER counts three kinds of error

The score comes from aligning the recognizer's output with a human reference transcript. Each mismatch is one of three errors:

• A substitution is a wrong word, such as "their" for "there".
• A deletion is a word the speaker said that the system left out.
• An insertion is a word the system added that the speaker did not say.

The total of these errors is divided by the number of words in the reference. The formula is WER = (S + D + I) / N, where S, D and I are the substitutions, deletions and insertions, and N is the number of words in the correct transcript. The denominator is the word count of the reference ². Because insertions are counted, the score can exceed 100 percent ³. A system that outputs two wrong words for every input word scores 200 percent, even though it communicates nothing ⁴.

A concrete example

Take the reference "the meeting starts at nine" and the output "the meeting starts at nine thirty". The output has one insertion and no other errors. Five reference words means WER = 1 / 5 = 20 percent. A dictation tool that drops a word in ten sentences would have a much lower score than one that rewrites every sentence, even if both sound fluent. WER counts literal differences, not meaning. One study notes that it measures only literal correctness, not semantic correctness ⁵.

How the number is reported

Published WER figures depend on text normalization. The Open ASR Leaderboard removes punctuation, casing and disfluencies such as fillers before scoring the normalized transcripts ⁶. A score on normalized text is not the same as a score on raw output. When you compare two results, check whether both used the same normalization and the same test set.

WER also does not match what people judge as useful. In a 2017 study of caption usability, WER correlated only 0.11 with the same ratings ⁷. A low WER is necessary for good dictation, but it does not guarantee that the text is easy to read or correct.

Why it matters for dictation

For dictation, WER tells you how much you will need to fix. A 2025 test of a single podcast episode found a WER of 10.3 percent for Apple's speech API, against 1.5 percent for Whisper Large V3 Turbo ^{8 9}. That single episode is not a general ranking. The most useful test is to measure WER on recordings that sound like your own speech, microphone and vocabulary. Our guide on word error rate, how to calculate and read it shows the steps.

Related terms

• Automatic speech recognition (ASR): the task that WER scores.
• Real-time factor (RTF): the speed counterpart to accuracy.
• Inverse text normalization (ITN): the formatting step that turns spoken numbers into digits, which affects scores.
• Speaker diarization: labeling who said each word, which needs its own error measure.

Guides that cover it

• Word error rate (WER): how to calculate and read it
• How accurate is speech recognition? The real numbers
• Speech recognition benchmarks and leaderboards