Measure Transcription Accuracy With WER
Paste the correct transcript in one box and the transcript you want to score in the other, then read the Word Error Rate. You also get the character error rate, the word accuracy, and a word-by-word diff that marks every substitution, deletion, and insertion. It all runs in the browser, so neither transcript is uploaded anywhere.
WER is the standard way to put a number on how close a transcript is to the truth. It is what speech-recognition papers report, what teams use to compare one transcription tool against another, and what you would use to decide whether a transcript is clean enough to ship or needs another pass.
What Word Error Rate Actually Measures
WER counts the edits it takes to turn the transcript you are checking into the reference, then divides by the length of the reference:
WER = (S + D + I) / N
S is substitutions, a word that came out wrong. D is deletions, a reference word the transcript dropped. I is insertions, an extra word the transcript added. N is the total number of words in the reference. The tool finds the smallest set of those edits by aligning the two word sequences, the same edit-distance idea behind spell-checkers, so the count is the fairest possible reading rather than a naive left-to-right compare.
A WER of 0 means the two match word for word. A WER of 0.10 means one edit for every ten reference words. It can go above 1.0, which surprises people the first time: if the transcript is padded with far more words than the reference, the insertions alone can outnumber the reference length, so 120 percent WER is a real result, not a bug.
How to Calculate WER
- Paste the reference, the transcript you know is correct, into the left box. This is what everything is measured against, so it needs to be the trustworthy one.
- Paste the transcript you want to score into the right box.
- Choose whether to ignore capitalization and punctuation, then read the WER, CER, accuracy, and the marked-up diff below.
The diff is the part worth reading. A single number tells you how bad, the diff tells you where and why. You can see at a glance whether the errors are scattered mishearings or one whole dropped sentence, and that changes what you do next.
Why Your WER Might Look Worse Than the Transcript Is
Raw WER is unforgiving about things that are not really errors. “Hello.” and “hello” count as a substitution if you leave punctuation and case in. So do “OK” versus “okay”, or “twenty” versus “20”. None of those change the meaning, but each one adds to the count.
That is why the two toggles matter. Ignoring capitalization and punctuation is the normal setting for comparing spoken content, because you care whether the words are right, not whether a comma landed. Turn them off when formatting is part of what you are grading, for example a subtitle file where casing and punctuation are the deliverable. There is no single correct setting. The honest move is to pick one, keep it the same across every transcript you compare, and say which one you used.
Numbers, spelled-out words, and contractions are the other common inflaters. If your reference writes “cannot” and the transcript writes “can’t”, that is a substitution even though a listener would never notice. For a strict benchmark you normalize both sides first; for a quick sanity check the toggles are usually enough.
WER, CER, and Word Accuracy
Word accuracy is just 1 - WER, flipped so higher is better. It is the friendlier number for a report, though it carries the same caveats, and it is clamped at zero because a transcript can be worse than “all words wrong” once insertions pile up.
Character Error Rate does the same edit-distance count on characters instead of words. CER is the better metric when word boundaries are unreliable, for languages that do not space words the way English does, or when you want partial credit for a word that is nearly right. “recognize” versus “recognise” is a full substitution under WER but a single-character edit under CER, so CER reads lower and often fairer for close misses.
Use WER as the headline for English speech, glance at CER when a low WER still feels too harsh on near-misses, and quote word accuracy when you are handing a result to someone who does not live in error rates.
Who Checks Word Error Rate
Teams comparing transcription vendors run the same audio through each one, transcribe a clip by hand as the reference, and score every tool’s output against it. WER turns “this one felt better” into a number they can defend.
Researchers and students working on speech recognition report WER because it is the shared yardstick. A model change is only an improvement if the WER drops on a held-out reference set.
Captioning and localization teams check WER before a transcript goes to translation or to a subtitle file, since an error at this stage gets multiplied downstream. And anyone who has paid for transcription and wants to know if they got what they paid for can measure it here instead of guessing.
After You Have the Number
If the transcript you are scoring came out rough and you want a cleaner one to start from, the transcript generator produces the transcript with speaker labels, and the AI transcript checker proofreads a single transcript when you do not have a reference to measure against. To turn a scored transcript into a subtitle file once you are happy with it, the text to SRT converter handles the timing.
Recordings are transcribed by Whisper Large-v3, and you can see the models and setup behind the accuracy on the accuracy page.
FAQ
What is a good Word Error Rate?
It depends entirely on the audio. Clean, single-speaker read audio can land in the low single digits, while noisy, multi-speaker, or heavily accented recordings run much higher for every tool. There is no universal pass mark. Compare tools on the same audio with the same reference, and judge the WER against that, not against a number from a different dataset.
How is WER calculated?
By aligning the transcript to the reference to find the fewest edits between them, counting substitutions, deletions, and insertions, and dividing that total by the number of words in the reference. This tool does the alignment for you and shows each edit in the diff.
Can Word Error Rate be more than 100%?
Yes. If the transcript adds many more words than the reference has, the insertions alone can push the total edits past the reference length, so WER goes above 1.0. A very short reference against a long, wrong transcript is the usual way to see it.
Does capitalization and punctuation count?
Only if you leave those toggles on. By default the tool ignores both, because for spoken content most people care whether the words are right, not the casing. Turn them on when the formatting itself is what you are grading.
What is the difference between WER and CER?
WER counts errors in whole words, CER counts them in characters. CER gives partial credit for a word that is almost right and works better for languages that do not separate words with spaces. For English speech, WER is the usual headline and CER is a useful second look.
Is my transcript private?
Yes. The whole calculation happens in your browser with JavaScript. Neither the reference nor the transcript you are scoring is sent to a server, so nothing is uploaded, stored, or logged.
How do I measure my transcription tool’s accuracy?
Transcribe a clip by hand, or correct one transcript carefully, and use that as the reference. Run the same audio through the tool you are testing and paste its output as the transcript to score. The WER is that tool’s error rate on that clip. Use several clips for a reading you can trust.
