What is word error rate in transcription?
Quick answer
Word error rate, or WER, is the standard measure of how accurate a transcript is. You compare it with what was really said and count the words that were swapped, dropped or added, then divide by the number of words spoken. Lower is better, and zero means perfect.
Last updated
Take the free module →Module 31 of the free AI Apps Worth Knowing course, about 10 minutes. No account needed.
Updated · By Robert Breen
Why it matters for a small business
Vendors advertise accuracy as one big percentage, and it hides the cost that matters. A transcript can be 99% right and still swap the one word that changes a client quote, a dollar figure or a name. If that quote goes into a report or a proposal, the error now carries your firm's name.
Knowing how WER works tells you what to check. Errors cluster in names, numbers, jargon and short words that sound alike. So you do not need to re-listen to a whole call. You play the lines you plan to quote or act on, and you compare a short passage against the source when you first try a tool. That spot check takes minutes and catches the mistakes that cost the most.
Working out WER on one sentence
| Item | Value |
|---|---|
| What was said | Even a short email would have been fine (8 words) |
| What the transcript wrote | Even a short email would have been fun |
| Substitutions, deletions, insertions | 1, 0, 0 |
| WER | 1 divided by 8, or 12.5% for this sentence |
One wrong word in a long call is a tiny WER, and can still flip the meaning of a quote.
In a real lesson: Speak AI: Turn Customer Interviews into Themes and Quotes
The free Speak AI lesson comes with both halves you need to measure accuracy: the 2:58 Harbor Lane Accounting interview as an mp3, and the script as a .txt file to compare with the transcript. Insights reports 425 words and says transcription was confident throughout.
Even so, the last chapter catches a mistake. At 00:59 Mike says a short email would have been "fine", and the transcript wrote "fun". The lesson tells you to press play on that line before you quote it. It is the same habit the Wispr Flow lesson teaches for dictation: read names, days and numbers before you send.
Try this lesson free or read the step-by-step guide.
Common confusions
Word error rate vs confidence
Confidence is the tool's own guess about how sure it is. WER is measured against what was really said. A confident transcript can still have errors.
Word error rate vs speech-to-text
Speech-to-text is the technology. WER is the score you give its output on a particular recording.
Tips
- Test a new tool on a recording you have a script for, and compare a paragraph.
- Play every line you plan to quote, especially names and numbers.
- Better audio lowers WER more than any setting: one voice at a time, close to the microphone.
Related terms
More Voice agents terms
Where you use it: free lessons
- Speak AI: Turn Customer Interviews into Themes and Quotes (Speak AI, 10 min)
- Wispr Flow: Answer Emails and Write AI Prompts by Voice (Wispr Flow, 10 min)
AI tools that use it
Frequently asked questions
- What is a good word error rate?
- There is no single cutoff. Clear audio with one speaker can come out with very few errors, while noisy calls with jargon do worse. Judge by whether the words you rely on are right.
- How do I check transcription accuracy without a script?
- Pick a short passage, play it, and correct it by hand. Count the fixes and divide by the words in that passage.
- Why does a transcript get names wrong?
- Names and firm terms are rare words the model has seen less often. Dictionary or vocabulary features in some tools help with that.