What Is Word Error Rate? Speech Accuracy Explained
Word error rate counts substitutions, insertions and deletions against a reference transcript. Here is how to read the number, and the four ways it misleads you.
Word error rate is the standard accuracy score for speech recognition. It is the number of word mistakes a transcript contains divided by the number of words in the correct version, expressed as a percentage. A 5% word error rate means one word in twenty is wrong. The formula counts three kinds of mistake, and understanding which kind you are getting matters more than the headline number.
The word error rate formula
WER = (Substitutions + Insertions + Deletions) / Words in the reference
A **substitution** is a wrong word in the right place. The speaker said "invoice", the transcript says "in voice".
An **insertion** is a word that was never said. Background noise becomes "the", or a cough becomes "uh-huh".
A **deletion** is a word that vanished. The speaker said "do not send it", the transcript says "do send it".
Worked example. Reference: "please send the invoice to Priya on Friday", nine words. Transcript: "please send invoice to Prya on Friday and". One deletion (the), one substitution (Prya for Priya), one insertion (and). Three errors over nine reference words gives a WER of 33%.
Notice that the three errors are not equally bad. Losing "the" costs nothing. Misspelling a name is annoying. And in a slightly different sentence, a deleted "not" would have inverted the meaning entirely. WER treats all three as one error each.
Why word error rate can exceed 100%
This surprises people the first time they see it, and it is not a bug in the metric.
The denominator is the length of the correct transcript, but insertions have no upper bound. A model that hallucinates thirty words of output for a five-word clip has scored more errors than there were words to get right. Any WER above 100% almost always means the system is generating text from noise rather than transcribing.
If you ever see this in your own evaluation, stop looking at the score and go listen to the audio. It is usually silence, a very low signal level, or a codec mismatch that turned your file into static.
What a good WER looks like in 2026
Context matters more than the number, but rough bands:
WER | What it feels like in a product |
|---|---|
Under 5% | Reads as correct. Occasional name or acronym is wrong. |
5 to 10% | Clearly usable, obviously machine-made. Needs a human glance before it is sent anywhere. |
10 to 20% | Meaning survives, individual sentences do not. Fine for search, bad for records. |
Over 20% | You are reconstructing rather than reading. |
Vendors publish numbers at the low end of that table. Google reported an average 2.6% WER for pre-recorded audio on its Gemini 3.5 Transcribe model, and 4.0% for live streaming, figures we looked at in more detail here. Those are averages across more than 85 languages, and averages hide the spread.
The four ways a WER number misleads you
**It is measured on a benchmark, not on your audio.** Benchmark corpora are recorded deliberately: reasonable microphones, cooperative speakers, limited crosstalk. Your users are in a car, or a cafe, or on a laptop microphone eighteen inches away. The same model can be three or four times worse on real input than on the number in the launch post. The same model's own FLEURS benchmark score was 5.04% against a 2.6% headline, which tells you how much the choice of test set moves the result.
**It averages across languages that behave nothing alike.** A model advertised at 3% across 85 languages might be at 1.5% in English and 12% in a language with fewer training hours. If your users are not speaking the language the number was optimised for, the number is not about you.
**It weights every word equally.** "The" and "not" cost the same. So do "£1,500" and "£15,000". For most business uses, a small number of high-stakes tokens carry nearly all the value, and WER cannot see them. This is why a 4% WER transcript can be perfectly readable and still get a price wrong.
**It says nothing about formatting.** Punctuation, capitalisation, and speaker labels are usually excluded from the calculation entirely. A transcript can score beautifully and still be a wall of lowercase text with no sentence breaks.
How to measure WER on your own audio
The vendor's number is a filter for shortlisting, not a decision. Measuring your own takes an afternoon.
Collect 20 to 30 real clips from your actual users, covering your worst conditions rather than your best. Get consent first.
Transcribe them by hand. This is the tedious part and there is no shortcut. These are your reference transcripts.
Run the same clips through each candidate provider.
Score with a standard tool. `jiwer` in Python is the common choice and is three lines of code.
Read the errors, not just the score. Which words are being lost? If it is your product names and your customers' surnames, the fix is a vocabulary hint list, not a different provider.
That last step is where the value is. Two providers at the same WER can fail completely differently, and one of those failure patterns is usually much cheaper for you to work around.
Related measures worth knowing
**Character error rate (CER)** does the same arithmetic on characters instead of words. It is the standard for languages without clear word boundaries, and it is more forgiving of near-misses: "Priya" against "Prya" is one whole word error but only one character error out of five.
**Diarisation error rate** measures who-said-what, not what-was-said. If you are transcribing meetings or calls with several speakers, this is the number that determines whether the output is usable, and it is completely independent of WER.
WER is also a distant relative of the evaluation metrics used for language models generally. If you want the parallel concept for text generation, perplexity is the closest analogue, and it comes with a similar warning about being easy to over-trust.
Where this fits
Speech recognition is one of the older pieces of applied machine learning, and its evaluation habits predate the current model generation by decades. That is why WER is a strict, mechanical, meaning-blind metric while newer benchmarks try to score usefulness. Both approaches have the same underlying problem, described in how AI models work: a single number cannot tell you whether a system will hold up on your particular input.
If you are wiring speech into something you have built, the practical follow-on is wiring speech capture into a product, which covers the parts of the job that no accuracy score predicts.
FAQ
Is a lower word error rate always better?
Lower is better for the same test set. Comparing WER figures measured on different audio tells you almost nothing, which is why vendor-to-vendor comparisons only mean something when both were run on the same clips.
What is a good WER for a phone call?
Telephone audio is narrowband and noisy, so expect meaningfully worse results than studio recordings of the same speakers. Judging call transcription against a benchmark recorded on good microphones will make every provider look broken.
Does WER include punctuation errors?
Usually not. Standard WER scoring normalises the text first, stripping punctuation and case. A model can therefore score well and still produce output that needs formatting work before a human would want to read it.
Can I improve WER without changing provider?
Often, yes. Better microphone input, a vocabulary hint list containing your domain terms, and choosing the batch endpoint over the streaming one all reduce errors without switching vendors.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
