Glossary: Meetings

What is Word error rate?

A score for transcription accuracy: the share of words that were wrong, missing or added, compared with a perfect transcript.

Updated Nov 1, 2026By Bruce W.

Word error rate, or WER, is the standard way to score speech-to-text. Compare the machine’s transcript with a perfect one and count three kinds of mistakes:

  • Substitutions: a wrong word (“cell” for “sell”).
  • Deletions: a missed word.
  • Insertions: an extra word.

Add them up and divide by the number of words that were actually said. 5 mistakes in 100 words is a WER of 5%.

What it doesn’t tell you

  • Which words were wrong. Getting a name or a number wrong matters more than a missed “um”.
  • Real-world sound. Tests use clean recordings. Your meetings have laptop mics and people talking over each other.
  • Punctuation and speakers. Most tests ignore them.

A worked example

What was said (12 words): “Let’s move the launch to the twenty first and tell the team.”

What the transcript says: “Let’s move the lunch to the twenty first and tell team.”

  • Substitution: “lunch” for “launch”.
  • Deletion: the “the” before “team” is missing.

That’s 2 mistakes in 12 words, a WER of about 17%. One mistake changes the meaning. The other barely matters. WER counts them the same.

Getting a lower WER on your own calls

  • Use a headset or earbuds with a mic instead of a laptop mic across the room.
  • Record your computer’s sound directly rather than through speakers.
  • Pick a bigger model if your computer can run it. See speech-to-text model.
  • Cut background noise: close the window, and mute when you’re not talking.

The biggest gains usually come from the microphone, not the software.

In Fetch

Settings > Transcription in Fetch lists each model’s Mistakes: wrong words per 100 on a public test. That’s WER, written as a number you can picture. See Choose a speech model.

Questions people ask

What's a good word error rate?

Under about 5% reads as very good for clear speech. Real meetings with crosstalk and accents score worse than public tests.

Why is my word error rate worse than the published number?

Published numbers come from clean test recordings. Real calls have laptop mics, accents, jargon and people talking over each other, which all add mistakes. Testing on your own recordings gives a truer picture.

Still scratching your head?

Send us a note and a real person will get back to you.

Contact support