Word error rate
Word error rate (WER) is the standard measure of speech-recognition accuracy. It is the number of substituted, deleted and inserted words in a transcript, divided by the number of words in the correct reference.
Also called: WER, transcription accuracy
Take a recording, transcribe it by hand to get the reference, run the recognizer, and count the differences. Every word the machine got wrong (substitution), left out (deletion) or added (insertion) is an error; the total divided by the reference word count is the rate. A WER of 0.10 means roughly one word in ten is wrong.
The figure depends heavily on the audio it was measured on. Clean studio speech, telephone audio, accented speech and noisy environments give very different results for the same system, so a quoted rate without its test conditions says little. Errors are also not equal: mishearing "the" costs nothing, mishearing a house number costs a missed appointment.
For a buyer, WER is the reason to test with real callers and real vocabulary rather than trust a headline accuracy figure, and to ask how the agent behaves when it is unsure of a word, confirming a spelling is cheap; acting on a wrong one is not.