Can voice recognition understand dysarthric speech?

Yes for mild dysarthria, and far less reliably for severe dysarthria. In a 2026 test, the leading commercial services made about 1 to 2 percent word errors on mild dysarthric speech, but every system exceeded 51 percent on severe speech ^{1 2}. In the studies below, training a recognizer on one speaker's own recordings produced the largest gain for severe speech.

Dysarthria is a speech disorder with a physical cause

Dysarthria is a speech disorder that happens when brain or nerve damage changes the way the muscles used for speech work ³. It can follow Parkinson's disease, stroke and other neurological conditions, and its effect on speech ranges from slight slurring to speech that is hard for listeners to follow. That range is why the recognition results below differ so much from one speaker to the next.

Error rates rise sharply from mild to severe dysarthria

A 2026 study states that automatic speech recognition reaches word error rates below 5 percent on typical speech, while performance degrades sharply for dysarthric speakers ⁴. On mild dysarthria, the leading services reached about 1 to 2 percent word error. That figure comes from a small test with two mild speakers, so read it as a sign of what is possible, not a rate you should expect.

On severe dysarthria, every commercial system the study tested exceeded 51 percent word error ². The test used four severe speakers and default settings, and the study found that prompting changed the figures for some models. So the number shows what the systems did out of the box, not the best each one can do. Even so, at that error rate more than one word in two is wrong.

Personalizing a recognizer closes much of the gap for severe speech

The largest gain in the studies below came from training a recognizer on the speaker's own voice. In a Google study, personalized models cut the median word error rate on severely impaired speech from about 89 percent to 13 percent ⁵. That test used short home-automation phrases, so it does not predict the error rate for long dictation.

Google's 2019 study of speech from people with ALS reports 10 percent word error for mild dysarthria and 20 percent for more serious dysarthria after personalization ⁶. The same study found that 71 percent of the improvement came from five minutes of training speech ⁷. That is a share of the improvement, not an error rate, so do not read it as a word error figure.

Parkinson's speech shows a smaller but real gain. A baseline open-source recognizer made 36.3 percent word error on speech from people with Parkinson's disease, and fine-tuning on Speech Accessibility Project data brought that to 23.7 percent ⁸. The model was a research system, not a commercial product, and even after fine-tuning roughly one word in four was wrong.

It depends on severity, phrase length and the service you use

The result changes with four things. The first is severity: the mild and severe groups in the studies behave very differently, and a single average hides that. The second is the kind of speech. Several of the studies here used short or read phrases, so they measure that kind of speech, not free conversation or long dictation. The third is the service, because cloud systems change and a 2026 test describes one window in time. The fourth is the personal data: a personalized model needs recordings of your own speech, and the gain depends on how much you can supply.

When you compare figures, check whether they are absolute or relative. Some personalization results are reported as relative improvements, which can sound larger than the change in the error rate itself.

What to do

1. Test with your own words. Read a short passage of the phrases you actually use, record it and compare the transcript with what you said.
2. Use a personalized option if your device offers one. Expect to record your speech, and the research suggests a few minutes of training speech can go a long way.
3. Keep commands short. Short phrases are a practical test for any recognizer, and the personalization results for severe speech used short phrases.
4. Keep a backup way to enter text, such as a keyboard, a switch or another input your device supports, for the words the recognizer still misses.
5. Talk to a clinician if your speech has changed recently. A speech-language pathologist or your doctor can assess your speech and suggest tools that suit it. We do not give individual medical advice.

For the full comparison of research results and apps, read our guide to speech recognition for stuttering and dysarthria and the overview of accessibility and health. If you stutter rather than having dysarthria, see whether speech recognition works if you stutter. For which features support atypical speech today, see which speech recognition apps work for atypical speech. Our entry on dysarthria defines the term, and the word error rate entry explains how each percentage here is counted.