WHAT IT MEANS
The percentage of words the speech-to-text gets wrong. Lower is better.
Real example
10 wrong words out of 100 = 10% WER.
What’s normal
About 5–10% on clean English. 10–25% or more on phone audio in Indian languages.
The India angle
Numbers, names and addresses matter more than overall WER. One wrong digit in a policy number breaks the call.
Questions to ask your vendor
What’s your WER on 8kHz phone audio in Hindi?
How accurate are you on numbers, dates and names?
Was your WER measured on your data or ours?
Companies that do this
Other terms in Hearing (Speech-to-Text)
Supported by
