Updated October 2026 · Speech recognition and on-device transcription
Word error rate is the share of words in a transcript that are wrong, counting substitutions, deletions and insertions against a reference. It is useful for comparing tools on the same audio, but a single score hides which errors matter. Test on your own recordings and check the errors that affect meaning.
Word error rate, usually written as WER, is the most common way transcription accuracy is measured. It appears in research papers, product announcements and comparison charts. It is a useful measure, but it is easy to misread. A single percentage can hide the errors that matter most to you, and a tool that scores well on one kind of audio may do poorly on another.
This post explains what WER counts, why it can mislead, how to compare tools fairly and how to test a tool on your own material. It is written for creators, editors and anyone choosing a transcription tool.
WER compares a transcript with a reference transcript, which is a checked version of what was said. It counts the words that have been substituted, deleted or inserted, adds those up and divides by the number of words in the reference. A lower score means fewer errors. The measure is simple to calculate, which is one reason it is so widely used.
The result depends entirely on the reference. A careless reference, or one that treats spelling and punctuation differently, can change the score.
A WER figure treats every error as equal. A missed filler word counts the same as a wrong name, a wrong number or a reversed negative. For a creator, the errors that change meaning are far more serious than the ones that do not. Two tools with the same WER may produce very different transcripts for your purposes.
The audio also matters. A score measured on clean studio speech may not hold for a noisy phone recording, a fast-talking panel or a speaker with a strong accent. Scores quoted by vendors often come from test sets that do not match your work.
Some errors are more harmful than others. A negative that is dropped (“did not” becoming “did”) reverses a claim. A number that changes (“fifteen” becoming “fifty”) misleads. A name that is wrong damages trust. When you check a transcript, look first for these high-impact errors, not for the total score.
WER is useful for a rough comparison of tools on the same material, and for tracking whether a tool improves over time on your own audio. Use it as one measure among several. The time it takes to fix a transcript is often the most practical measure of all.
Vendors choose test sets, references and conditions that show their product well. A score from a vendor is a claim to be tested, not a fact to rely on. Ask what audio was used, how the reference was made and whether the test matches your work. An independent test on similar audio is more useful than any headline figure.
A score is most useful over time, on your own audio. If you run the same short test each month, with the same recording and the same reference, you can see whether a tool, a setting or your own recording habits are improving the result. Keep the reference transcript and the test file in a folder, so the comparison stays honest.
A small change in the score is not always meaningful, especially on a short recording. Look for consistent movement across several tests before drawing a conclusion.
When you report a transcription score to a client, a colleague or a funder, explain what it counts and what it does not. A plain statement, such as “about one word in twenty was wrong, and most of the errors were filler words”, is more honest and more useful than a percentage on its own. The people who read the report need the meaning, not only the number.
WER does not capture meaning, tone, timing or the usefulness of a transcript to a particular reader. Two transcripts with the same score can read very differently. Use WER as a starting point, and judge the transcript on what it does for your work.
A clean recording lowers the error rate for any tool, which makes comparisons easier to trust. The Vocal Checker shows clipping, noise and level problems in a recording you own. The Tunedzy Go page covers cleaning a phone recording. For the practical side of checking a transcript, read the interview accuracy post in this series.
Record voice-overs with live noise removal, or clean an existing file. Runs in your browser, nothing is uploaded.
Usually, but only on similar audio. Check the errors that change meaning as well.
Compare your transcript with a careful reference, count the substitutions, deletions and insertions, and divide by the reference words.
They use different test sets, references and conditions. Compare on your own audio.
Errors that reverse a claim, change a number or get a name wrong, because they mislead the reader.
