What Word Error Rate Really Means for Your Writing
Every dictation tool brags about accuracy, and the number they usually point to is word error rate, or WER. It is a genuinely useful metric, but it does not measure what most people think it does. Here is what WER actually tells you about your writing on a Mac, and what it quietly leaves out.
Key takeaways
- WER = (substitutions + insertions + deletions) divided by the number of words you spoke.
- Lower is better, and modern on-device models for Mac dictation reach low single digits on clear speech.
- WER measures faithfulness to your words, not readability, so it rewards keeping every "um" intact.
- What you actually want is a low WER plus AI cleanup, which is where BlaBlaType does the second half of the job.
What word error rate actually measures
Word error rate is the standard accuracy metric for any voice to text system. You take a perfect human transcript, called the reference, and compare it word by word against what the machine produced. Every place they disagree is counted as one of three error types: a substitution (the wrong word), an insertion (an extra word), or a deletion (a missing word). Add those up, divide by the number of words in the reference, and you have your WER.
So a WER of 5% means that for every hundred words you spoke, five were wrong in some way. A WER of 0% means a flawless match. The metric comes straight out of academic speech recognition research, including the original Whisper paper from OpenAI, which reports WER across dozens of datasets. It is the same yardstick that shapes the state of Mac dictation in 2026, because it lets very different models be ranked on a single number.
Why a lower WER matters for real writing
The practical payoff of a low WER is simple: you spend less time fixing the transcript. If your speech to text app misses one word in twenty, a paragraph of email needs a couple of quick corrections. If it misses one in five, you are effectively proofreading a stranger's rough notes, and the whole speed advantage of dictation evaporates. Since most people speak around three to four times faster than they type, the entire reason to dictate is to save that time, and a high WER hands it right back.
Accuracy also compounds with context. Names, jargon and product terms are the words a model is least likely to know, so they are the ones that drive your personal WER up. A custom dictionary that teaches the app your own vocabulary is one of the most effective ways to push the real-world number down, especially for technical work like when you dictate into Xcode on a Mac and half your sentences are variable names.
What WER does not tell you
Here is the catch that the marketing rarely mentions. WER measures how faithfully the system reproduces your exact words, filler and all. If you say "um, so, I think, you know, we should ship it," a perfect transcript keeps every one of those words. That is a WER of 0%, and it is also unreadable. The metric rewards transcription, not editing.
Real writing needs the second half of the job: punctuation, capitalization, paragraph breaks, and the quiet removal of "um" and "you know." None of that shows up in a WER score, because the reference transcript is just the raw stream of spoken words. This is why two apps with an identical WER can produce wildly different output, one a wall of run-on speech and the other a clean, finished paragraph. On-device AI cleanup, powered on Mac by Apple Intelligence, is what closes that gap. It fixes grammar, trims filler and adapts tone without touching a server. To understand why doing that locally matters, our piece on whether Mac dictation is private walks through where your words go, and privacy frameworks like GDPR are a big reason on-device processing is worth paying attention to.
WER across real conditions
A single WER number is really an average across many recording conditions, and it moves a lot depending on the situation. The table below shows the general pattern you can expect from a modern on-device model for voice to text on a Mac. These are directional, not lab-measured guarantees, and your own mileage depends on your microphone, accent and vocabulary.
| Condition | Effect on WER | On-device fix |
|---|---|---|
| Clear speech, quiet room | Lowest | Good mic, natural pace |
| Background noise | Higher | Voice activity detection |
| Names and jargon | Higher | Custom dictionary |
| Strong accent | Varies | Model choice, dictionary |
| Filler and false starts | No WER change | AI cleanup |
Notice the last row. Filler words do not raise WER at all, because a faithful transcript keeps them, yet they are exactly what makes raw dictation look messy. That is the clearest proof that WER alone is not the whole story for writing.
See low WER plus real cleanup on your Mac
BlaBlaType pairs on-device speech recognition with local AI cleanup, so you get accurate text that is already polished. No card needed for the trial.
Download for macOSHow to read accuracy claims in 2026
When a Mac dictation app quotes a WER, ask two questions. First, on what data? A 2% WER on a clean audiobook dataset is not the same as 2% on your noisy kitchen recording, so treat any single figure as a best case. Second, does the app also clean up the text, or does it stop at raw transcription? The best experience combines both: a strong on-device model to keep WER low, and local AI editing to make the result readable. You can see how the pieces fit together on the BlaBlaType overview, and compare tiers on the pricing page. WER is a good starting filter, just not the finish line.
Frequently asked questions
What is a good word error rate for dictation?
For clear speech in a quiet room, modern on-device models often land in the low single digits, meaning only a few words in a hundred are wrong. Anything in that range feels reliable for real writing, since a quick read-through catches the rest.
How is word error rate calculated?
Word error rate adds up the substitutions, insertions and deletions the system makes, then divides that total by the number of words in the reference text. A WER of 5% means five errors for every hundred spoken words.
Does a low word error rate mean the writing is good?
Not on its own. WER only measures how closely the transcript matches what you said, filler words included. Good writing also needs punctuation, formatting and cleanup, which is where on-device AI editing matters more than the raw number.