WER, short for Word Error Rate, is one of the most familiar measures used to evaluate automatic speech recognition systems. It answers a practical question: how closely does a transcript produced by a machine match the words that were actually spoken?
The calculation compares a machine-generated transcript with a reference transcript prepared by a person. It counts three kinds of differences: substitutions, where one word is recognized as another; deletions, where a spoken word is missing; and insertions, where the system adds a word that was never said. The standard formula is:
WER = (Substitutions + Deletions + Insertions) ÷ Number of words in the reference transcript
The result is usually expressed as a percentage. A transcript with no differences has a WER of 0%. A higher percentage means more word-level errors. In unusual cases, WER can exceed 100%, especially when a short reference sentence produces a much longer and inaccurate machine transcript.
Consider the sentence, “Please send the report today.” If a system produces, “Please sent the report,” it has made more than one type of mistake: “sent” substitutes for “send,” while “today” has been deleted. The exact score depends on the total number of words in the reference, but the example shows why simply counting visibly incorrect phrases is not enough.
WER is useful because it creates a common basis for comparing models, microphones, recording conditions, and languages. A speech recognition system may perform well in a quiet office but struggle in a busy café, during a phone call, or when several people speak at once. Accents, specialist vocabulary, background noise, fast speech, and poor microphone placement can all affect the result.
Still, WER should not be treated as a complete description of quality. Not every error has the same effect on meaning. Misrecognizing a person’s name, a medication, an address, or a number may be far more serious than missing a casual filler word. Conversely, a transcript can have a relatively high WER while remaining easy for a reader to understand if the errors involve repeated words or minor grammatical details.
The way a transcript is prepared also changes the score. Decisions about punctuation, capitalization, numbers, contractions, and spelling must be consistent between the reference and generated text. For example, “twenty four” and “24” may be treated as different unless the evaluation process normalizes them. This is why reliable testing usually defines its text-cleaning rules before calculating results.
For everyday users, WER is best viewed as a useful warning light rather than a final verdict. It can reveal whether a system is improving and where it fails, but human review remains important when accuracy carries real consequences. A lower score is valuable, yet the most suitable system is the one that performs reliably in the actual situations where people will use it.