WER (word error rate)

Word error rate counts the substitutions, deletions, and insertions a transcript needs to match a human reference, divided by the number of words in the reference. A WER of 11 means about one word in nine is wrong somewhere. It is the number every speech-recognition paper reports, which makes models comparable, and it hides everything about *which* words were missed.

That blind spot matters for meetings. Two models with the same WER can feel very different if one of them reliably drops names, numbers, and product terms, because those are the words people actually go back to look for. WER also says nothing about punctuation, capitalization, or whether the model wrote 'twenty twenty four' instead of '2024'.

Benchmark WER is measured on clean, read-aloud datasets. Real meeting audio has crosstalk, bad microphones, and accents, so expect the number you experience to be noticeably worse than any published figure.