Word error rate (WER)

Word error rate is the standard metric for scoring the output of an automatic speech recognition system against a human-verified reference transcript. It counts how many word-level edits — substitutions, deletions, and insertions — are needed to turn the system's output into the reference text, then normalizes by the number of words in the reference.

The formula is WER = (S + D + I) / N, where S, D, and I are the counts of substituted, deleted, and inserted words and N is the number of words in the reference transcript. A WER of 10% means, roughly, that one word in ten required an edit to match the reference; WER can exceed 100% if a system inserts far more words than the reference contains. This differs from precision and recall or f1 score, typically used for classification tasks, though similar edit-distance ideas appear in other text-comparison metrics.

Word error rate matters as the benchmark vendors and researchers use to compare ASR systems, and it directly affects downstream speech analytics quality, since errors compound through any analysis built on the transcript, including speaker diarization attribution. A common pitfall is comparing WER figures measured on different test sets or audio conditions as if directly comparable; WER is highly sensitive to accent, background noise, and domain vocabulary, so the number only means something alongside the conditions it was measured under.