Word Error Rate (WER)

Word error rate, abbreviated WER, measures speech-recognition accuracy by comparing a transcript with a reference. It counts substitutions, deletions, and insertions, then divides their total by the number of words in the reference. A lower score usually indicates a closer transcription, although WER does not capture every practical difference: punctuation, speaker labels, proper names, and a single critical number may matter more than several harmless wording changes. Scores also vary with accents, noise, microphones, vocabulary, and normalization rules. Teams therefore report WER alongside subgroup testing and task-specific measures that reflect whether users can successfully understand or act on the transcript.