Word Error Rate (WER): definition and how to read it
Word error rate (WER) is the standard accuracy score for speech recognition: the number of substituted, deleted and inserted words divided by the words actually spoken. Lower is better, and 0% means a perfect transcript.
Also known as: WER, transcription error rate
How it is calculated
Compare the recognizer's transcript with a human-checked reference. Count substitutions, deletions and insertions, add them up and divide by the number of words in the reference. If a caller says ten words and the system gets two wrong and drops one, WER is 30%. It can exceed 100% when many words are inserted.
Why a single WER misleads
WER treats every word equally, but 'the' and a loan amount are not equally important. It also depends hugely on the test set: clean read speech, noisy phone calls and mixed-language speech give very different results from the same model. A published figure is only meaningful with the corpus described, and a per-language number says more than an average.
Common mistake
Comparing vendors' WER claims as if they were measured the same way. Ask what audio, what languages, what normalisation of numbers and spelling, and whether the figure is production or benchmark. Better still, score your own calls.
Where this shows up at Vaami
Vaami reports under 6% word error rate across its production call corpus; per-language WER is not published.
See it on a real call.
Hear how a Vaami voice agent handles this on a live conversation in your language and your use case.