Word Error Rate Explained: How Accuracy Is Measured

When a dictation app claims to be "highly accurate," that claim usually rests on a single number: word error rate, or WER. It is the standard yardstick for speech to text, and once you understand how it is measured, you can read accuracy claims with a much clearer eye.

What word error rate actually measures

Word error rate compares two things: the transcript a speech-to-text system produced, and a reference transcript that is known to be correct. The reference is often called the ground truth. WER counts how many single-word edits it takes to turn the machine's output into the reference, then expresses that as a fraction of the reference length.

The idea comes straight out of computer science, where it is known as edit distance. Modern recognition models such as Whisper and Parakeet are still scored this way. If you want the technical background on how one of those models works, the Whisper speech recognition overview is a good starting point.

The three error types

Every mistake WER counts falls into one of three buckets. Understanding them is the whole game.

• Substitution. The system heard one word but wrote another. You said "their," it wrote "there." One substitution.
• Deletion. The system dropped a word that was actually spoken. You said "the quick brown fox," it wrote "the brown fox." One deletion.
• Insertion. The system added a word that was never spoken, often from background noise or a stray sound. You said "hello world," it wrote "hello there world." One insertion.

A good chunk of insertions and deletions trace back to how the audio was segmented before recognition ever ran. That segmentation step leans on voice activity detection, which decides which slices of audio contain speech. If that is new to you, we explain it in what voice activity detection is.

The WER formula, step by step

The formula is simpler than it looks:

WER = (Substitutions + Deletions + Insertions) / Words in the reference

Say your reference sentence has 10 words. The transcript has 1 substitution, 1 deletion, and 0 insertions. That is 2 errors divided by 10 reference words, which gives 0.2, or a 20% word error rate. Flip it around and you get a rough "word accuracy" of 80%, though accuracy and WER are not perfect mirrors once insertions push the error count above the reference length.

How dictation turns your voice into a scored transcript

WER only makes sense once you see where in the pipeline the errors appear. On a private, on-device app the whole path stays on your Mac: the microphone captures audio, a local model transcribes it, an optional AI cleanup step polishes it, and the text lands in whatever app you are typing into.

The distinction matters. WER is a fair test of the recognition step, where the goal is to match what you said word for word. Once the AI cleanup step removes filler, fixes punctuation, and tidies grammar, the wording deliberately diverges from your raw speech, so a raw WER against your exact words would punish improvements. BlaBlaType runs both steps entirely on your Mac, which also means you can benchmark accuracy without any audio leaving the device. That is central to the on-device privacy stack.

What counts as a "good" word error rate?

There is no universal pass mark, and this is where headline numbers get misleading. WER depends heavily on the audio. Clean speech recorded close to a good microphone in a quiet room will score far lower than the same words shouted across a noisy café. Accents, technical jargon, proper names, and overlapping speakers all push WER up regardless of how good the model is.

WER also treats every word as equally important, which real life does not. A transcript that drops an "um" and one that swaps a client's surname can post the same error count, yet only one of them ruins your document. That is why a custom dictionary for names and jargon can matter more to your finished text than a fractionally lower WER on a generic benchmark.

The practical takeaway: never trust a single advertised percentage. Compare tools on the same recordings, ideally your own, and weigh the errors that actually cost you time. If privacy is part of your decision, remember that any cloud benchmark had to upload audio to produce its number, whereas an on-device tool does not. We cover that trade-off in whether Mac dictation is private, and EU readers may also want the GDPR angle on dictation.

How to test word error rate yourself

You do not need a lab to get a real accuracy figure for your own voice. Take a paragraph you have already written, read it aloud, and dictate it with a tool like BlaBlaType. Then compare the transcript to your original, mark each substitution, deletion, and insertion, and divide by your word count. Do the same paragraph in a few conditions: quiet room, background music, your usual jargon. You will quickly learn far more than any advertised number can tell you, and you can weigh it against your budget on the pricing page. For a broader comparison of tools, our Superwhisper alternative roundup is a useful next read, as is Apple's own guide to using Dictation on Mac.