What is word error rate?
Checked against the product on October 22, 2026
Word error rate, or WER, is the standard way to measure how accurate a transcript is: it counts substituted, deleted and inserted words compared with a correct reference, divided by the number of words in the reference. It is a useful yardstick, but it does not tell you which errors matter.
Definition
To compute WER, a person writes the correct transcript of a recording. The automatic transcript is then compared with it one word at a time, counting three kinds of error: substitutions, where a wrong word replaces the right one; deletions, where a word is missing; and insertions, where an extra word appears. The total, divided by the number of words in the reference, is the word error rate. A rate of 5 percent means roughly one error in every twenty words.
An example
| Said | Transcribed | Error type |
|---|---|---|
| Book a flight to Lisbon | Book a flight to Lisbon | None |
| Book a flight to Lisbon | Look a flight to Lisbon | Substitution |
| Book a flight to Lisbon | Book flight to Lisbon | Deletion |
| Book a flight to Lisbon | Book a quick flight to Lisbon | Insertion |
Why a low WER is not the whole story
- All errors count the same, but a wrong brand name matters more than a missing 'the'.
- It measures words, not meaning: two transcripts with the same rate can differ in usefulness.
- Results depend heavily on the test audio, so a figure from one recording does not apply to yours.
- Punctuation and capitalisation are usually ignored, but viewers see them.
Using the idea in your own workflow
- Pick a sample clipChoose a minute of typical audio from your own recordings.
- Read alongCount the errors you can see in the captions.
- Note which kinds recurNames, numbers, specialised words and fast speech tend to repeat.
- Add corrections where it mattersFix recurring terms, and consider asking a guest to spell unusual names before recording.
Claims about accuracy
Be cautious with any tool that advertises a single accuracy percentage without saying how it was measured. Figures from clean test audio rarely match what you will get on a noisy podcast or a game stream. We do not claim a figure for Averoai.pro; the practical answer is to check the captions of each clip before you publish.
Frequently asked questions
Is a lower word error rate better?
Yes: fewer errors per word. But it should be read together with which errors they are.
Can WER be above 100 percent?
Yes, in theory, if there are many insertions. In practice it is usually well below.
Does WER include punctuation?
Usually not; it counts words. Punctuation errors still affect readability.
Does Averoai.pro publish its accuracy?
No figure is claimed. Quality depends on the audio, so review the captions of every clip.
Features that help with this
Related reading
- What is an A/B test?
An A/B test compares two versions that differ in one thing. Learn how to run a fair test with clips, why small samples mislead and what to conclude.
- What is an audio waveform?
A waveform is a picture of sound over time. Learn how to read it to find pauses, loud moments and clean cut points when trimming a clip.
- How to make clips from a video stored in Google Drive
Import a video stored in Google Drive and turn it into short clips: the sharing setting that makes the link work and the mistakes that make it fail.
- Dubbing vs subtitling: which suits your clips?
Dubbing replaces the voice; subtitling adds translated text. Compare cost, speed, feel and risks, and see which fits short clips for a new audience.
- What is an SRT file?
An SRT file is a plain-text subtitle format with numbered cues and timecodes. Learn how it is structured, where it is used and when you need one.
- What is forced alignment?
Forced alignment matches a known text to audio, word by word, to find exact timings. Learn when it helps, from corrected transcripts to translated captions.
Also available in Français