Does dictation work for people with Parkinson's disease?

Dictation works for most people with Parkinson's disease, but it makes more mistakes than it does for healthy people of a similar age. In our October 2026 test, Apple SpeechTranscriber had a word error rate of 13.81 percent ¹ on a read passage, Parakeet TDT v3 had 14.32 percent ², and Whisper large-v3-turbo had 21.84 percent ³. Word error rate is the share of words that are wrong, missing or added, so lower is better.

The same three engines scored 6.84 percent ⁴ (Whisper), 8.00 percent ⁵ (Parakeet) and 8.82 percent ⁶ (Apple) on healthy speakers aged 60 to 77 reading the same passage. Every engine made more errors on the Parkinson's group, but how much more depends on the engine and the text.

People with Parkinson's had higher error rates on every engine and both tasks

The test used the Italian Parkinson's Voice and Speech set. It holds readings by 24 people with Parkinson's disease aged 40 to 80 (some were recorded twice, so there are 52 passage readings), by 21 healthy speakers aged 60 to 77 and by 15 healthy speakers aged 19 to 29. We ran each recording through three engines on a Mac with an M2 Pro.

On the short passage, the lowest Parkinson's result (13.81 percent for Apple) is above the highest result for healthy speakers of a similar age (8.82 percent for Apple). On ten short sentences the pattern holds, but the gap is smaller:

Whisper had 7.47 percent ⁷ on the sentences for people with Parkinson's against 4.04 percent ⁸ for healthy older speakers. Parakeet had 7.62 percent ⁹ against 5.30 percent ¹⁰. Apple had 8.27 percent ¹¹ against 5.30 percent ¹².

Only some of the gaps are beyond chance

With 24 people in the group, a gap can come from who happened to be recorded. We compared each Parkinson's result with the healthy speakers aged 60 to 77 and allowed for the test's 9 comparisons when building the 95% intervals.

• Passage, Whisper and Parakeet: the interval excludes zero, so the extra errors are a measured difference. For Parakeet it only just excludes zero.
• Passage, Apple: the interval includes zero, though only barely. The test cannot tell this gap apart from chance.
• Short sentences, all three engines: every interval includes zero. The sentence results are higher for Parkinson's, but we cannot call the gap real.

The intervals are wide. A result that "includes zero" is not proof of no difference: with this few people, a modest gap would not show up. Individual speakers also vary a great deal. A few people with very high error rates pull each average up, and the worst single speaker on the passage had an error rate several times the group average on every engine.

Whisper's passage gap to healthy speakers is the widest, yet the engines are level within the Parkinson's group

Whisper's error rate on the passage for people with Parkinson's (21.84 percent ³) is several points above Apple's and Parakeet's, and its gap to healthy older speakers is the widest of the three in points. Within the Parkinson's group, though, no engine difference is measured on either task: the three count as level, because the interval allowing for the engine comparisons includes zero, and Whisper's errors sit in a few recordings ^{13 14 15}. On the short sentences the three engines also landed within a narrow band ^{16 17 18}.

The only engine differences we measured are on the passage. For young healthy speakers Parakeet made fewer errors than Apple and than Whisper ^{19 20}, and for healthy speakers aged 60 to 77 Whisper made fewer errors than Apple ²¹. We compare the engines in more detail in which engine understands Parkinson's speech best.

What the numbers do and do not measure

• The speakers read fixed text aloud. Dictating your own words, with pauses and corrections, may give different results.
• All speakers are Italian, from one region, and the engines were run in Italian. English or other languages may differ.
• The Parkinson's group is aged 40 to 80 and the comparison group 60 to 77, so ages are similar but not identical. We have no measure of how advanced each person's disease was, and more advanced Parkinson's may fare differently.
• We tested only on-device engines on a Mac, not cloud services.
• The set does not tell us about recording conditions, so we cannot say how much a microphone changes the result.

What helps if you dictate with Parkinson's

1. Pick the engine for your kind of text. On a longer passage Whisper's average error rate was higher for this group, but the test measures no difference between the engines, so treat the three as level. Try more than one on your own voice.
2. Use a close microphone. Our data does not test this, but Apple advises that a headset microphone may help in a noisy room or one with excessive echo ²².
3. Check the text. Even the lowest passage result in our test is a word error rate of 13.81 percent for the group, and some individuals fared far worse. Read dictated text back before you send it.
4. Keep a keyboard or another method as a backup for days when your voice is weaker.

Dysarthria, slurred or weak speech from muscle control problems, affects an estimated 44 to 88 percent of people with Parkinson's disease ²³, so it is worth testing a tool on your own voice before relying on it. This page describes test results and is not medical advice.

See the full results in the full dataset. For other speech differences, read our guide to speech recognition for speech impairments and the wider accessibility overview. Related data: how engines handle accents and an on-device speech recognition benchmark. If you also wonder about age alone, see whether speech recognition is less accurate for older adults.