RexaThe short answer
WER is the standard accuracy metric for speech recognition, and the standard way vendors are compared. It is genuinely useful and routinely over-read: a single percentage cannot tell you whether the errors fell on filler words or on the account number. For voice AI, WER is a screening metric, not an acceptance criterion.
In detail
WER is computed by aligning the hypothesis against a reference transcript and counting the minimum edits needed to reconcile them: substitutions, deletions and insertions, divided by the number of words in the reference. Because insertions are counted but not bounded by the reference length, WER can legitimately exceed 100% when a recogniser hallucinates text into silence.
The metric weights every word identically, which is exactly wrong for transactional calls. Missing "um" and missing "not" cost the same on paper. So does confusing "fifteen" with "fifty" in a payment amount. Teams that care about outcomes therefore track entity error rate alongside WER — accuracy on the specific fields the call exists to capture, such as dates, amounts, postcodes and names.
WER numbers are only comparable when the test conditions are. Audio bandwidth, codec, accent distribution, background noise, and the text-normalisation rules used for scoring all move the number by more than the differences between competing models. A vendor benchmark on clean wideband read speech tells you very little about performance on 8 kHz calls from mobile phones in cars.
Related terms
Automatic speech recognition
Automatic speech recognition is the conversion of spoken audio into written text by a machine.
Speaker diarization
Speaker diarization is the partitioning of an audio recording into segments labelled by who was speaking.
Hallucination
A hallucination is model output that is fluent and confident but not supported by the model’s inputs or by fact.
Structured outcome
A structured outcome is a machine-readable classification of what a call achieved, drawn from a defined set of possible results.