What is a good word error rate?

Below 5 percent is the benchmark for typical speech, and harder audio sets a higher bar

If you want one number, use 5 percent for typical speech. A 2026 study of commercial recognizers notes that automatic speech recognition reaches word error rates below 5 percent on typical speech ¹. On clean recordings of ordinary speakers, a score well above that points to a problem in the setup: the audio, the microphone, or how well the model fits the voice.

The benchmark is a starting point, not a verdict. A 5 percent word error rate (WER) means about one word in twenty is wrong, and two systems with the same score can make very different mistakes. What counts as good depends on three things: the audio, the speaker and what a wrong word costs you. The sections below take each in turn. For the arithmetic, see how to calculate word error rate, and for the full picture, our guide to word error rate, part of our overview of speech recognition.

Human transcribers set a floor, so even the best scores are rarely zero

Reference transcripts are written by people, and people make errors. In a 2017 IBM study, the best of three human transcribers reached a word error rate of 5.1 percent on the Switchboard telephone test, after quality checks ². On the harder CallHome telephone test, the best human reached 6.8 percent ³. Those are old telephone conversations, not dictation, but they show that a score near zero is not a realistic goal on hard audio. Human and AI transcription compares the two in more detail.

Your speech and your audio move the target by a wide margin

The same metric can give very different scores for different speakers and recordings. The table gathers the figures from the studies above.

Dysarthria is a motor speech disorder, and it is the clearest case. For someone with severe dysarthria, the target is set by what that person can work with, not by the typical-speech benchmark. Even after fine-tuning on accessibility data, roughly one word in four was wrong in the Parkinson's study ⁶. Read the score as a claim about one speaker group, not about speech in general.

A dictation study points to about 10 percent as a usable ceiling

In a 2004 study, high school students with and without learning disabilities learned to use speech recognition with total word errors under 10 percent ⁷. That is an upper bound from one study of students learning to dictate, not a standard for every writer. The study did not measure time saved, so treat 10 percent as a rough line for first drafts, not as a rule.

A low score does not guarantee readable text

WER counts every error the same way. A dropped "um" and a wrong surname each add one error, and the score cannot tell them apart. In a 2017 study of automatic captions, word error rate correlated only 0.11 with the usability ratings of deaf and hard of hearing users ⁸. A good score on paper can still leave a transcript that is hard to read or wrong where it matters. The WER guide explains why the number hides these differences.

How to decide what good means for you

1. Name the words that must be right. Names, numbers, doses and dates are common examples. Count them separately, because WER gives them no extra weight.
2. Record a short sample of your own. Use your microphone, your room and your usual speaking style, then transcribe it by hand.
3. Score every system the same way. Use one set of text rules for the reference and for each output. A different normalizer can change the result, so a comparison across rules is not fair.
4. Set the bar before you test. Decide the acceptable score from the cost of an error, so you do not pick a threshold after seeing the results.

For a first transcript to score, the free audio to text tool runs in your browser. Yap, which we make, is an on-device dictation app for Mac and Windows, so you can compare its output on your own voice with your hand transcript. For the mechanics of the score, see the question on how to calculate it, the question on whether it can exceed 100 percent and the question on WER in Python.