Word error rate (WER)
Also: wer
The standard measure of speech-to-text accuracy: the share of words a transcription gets wrong. Lower is better, so unlike most benchmarks it is a score you want to minimise.
Word error rate is how speech recognition is scored. It counts the words a model gets wrong, as substitutions, insertions, and deletions, against a reference transcript, and expresses that as a percentage. A WER of 5% means roughly one word in twenty is off.
The key thing to remember is that lower is better, the opposite of most benchmarks, so a model at the top of a speech leaderboard has the smallest number. It is the figure quoted for models like Whisper, usually as an average across several test sets, since accuracy varies a lot with accent, noise, and domain.