Dictation Accuracy With Parkinson's Speech, Tested

Speech recognition makes more mistakes on people with Parkinson's disease than on healthy speakers of a similar age, but only on a longer read passage, and only for two of the three engines we tested. Whisper large-v3-turbo made 21.84 percent word errors on people with Parkinson's ¹ against 6.84 percent on healthy speakers aged 60 to 77 ². Parakeet TDT v3 made 14.32 percent against 8.00 percent ^{3 4}. For Apple SpeechTranscriber the difference is not distinguishable from chance, and neither is any difference on short sentences.

Our data has 15 data points ⁵. All come from the Italian Parkinson's Voice and Speech set (University of Bari and ASL Bari): 28 recording sessions of people with Parkinson's aged 40 to 80 (about 24 people), 21 healthy speakers aged 60 to 77 and 15 healthy speakers aged 19 to 29. We ran three engines on a Mac with an M2 Pro in October 2026. Lower numbers are better.

Two of three engines make measurably more errors on a read passage

The passage was read twice by every speaker. Here are the six figures for the two groups that matter for the question.

For Whisper, the Parkinson's figure of 21.84 percent ¹ sits well above the older healthy figure of 6.84 percent ². The 95% interval for that difference, adjusted for the number of comparisons in the test, excludes zero, so it is a measured difference. It is also the widest gap of the three engines in points.

For Parakeet, the figures are 14.32 percent ³ and 8.00 percent ⁴. That interval also excludes zero, but only just, so treat the size of the gap as uncertain.

For Apple, the figures are 13.81 percent ⁶ and 8.82 percent ⁷. The Parkinson's figure is higher, but the interval for the difference includes zero. With this many speakers, we cannot say that Apple's gap is real, and we cannot say it is absent.

Short sentences show no measured gap for any engine

Ten short sentences were read by the older healthy speakers and the people with Parkinson's only. On these, every engine scored higher for Parkinson's, and in every case the interval for the difference includes zero.

The people with Parkinson's also scored lower error rates on the sentences than on the passage for all three engines (for Whisper, 7.47 percent against 21.84 percent). Two things differ between the tasks, so the data cannot say which one matters: the passage is longer, and the sentence rows cover 22 speakers with 26 recordings while the passage rows cover 24 speakers with 52. The sentence figures are therefore not a clean test of the passage figures.

Age explains part of the gap, so compare with healthy speakers of the same age

Older healthy speakers are harder for every engine than young ones. On the passage, young speakers scored 1.79 percent with Parakeet ¹⁴, 3.04 percent with Whisper ¹⁵ and 3.89 percent with Apple ¹⁶. The older healthy speakers scored 8.00, 6.84 and 8.82 percent for the same three engines. All three age gaps are measured differences.

For Parakeet, the age gap is about as large in points as its Parkinson's gap against older healthy speakers. Comparing people with Parkinson's with young adults would therefore have overstated the effect of the disease, which is why the research question uses speakers of a similar age.

The engines are level for people with Parkinson's, even where the figures look far apart

In the Parkinson's group the passage figures run from 13.81 percent for Apple ⁶ and 14.32 percent for Parakeet ³ to 21.84 percent for Whisper ¹. That looks like a real ranking, but the comparison of the engines on the same recordings does not support one ¹⁷. Whisper's errors sit in a few recordings, so its average is pulled up by a handful of them. On short sentences the three Parkinson's figures (7.62, 7.47 and 8.27 percent) are level too.

Engine differences are measured only on the passage, and only for healthy speakers:

• Young speakers: Parakeet made fewer errors than Apple and than Whisper (1.79 percent against 3.89 and 3.04 percent).
• Older healthy speakers: Whisper made fewer errors than Apple (6.84 percent against 8.82 percent). Parakeet, at 8.00 percent, is level with both.

The order therefore changes between groups, and none of these orderings is measured for people with Parkinson's.

Averages hide a few recordings that go badly wrong

The averages are driven by a minority of recordings. In every Parkinson's row, the typical speaker scored below the group average, while the worst single recording was far above it. The same pattern appears, less strongly, among the older healthy speakers. A group average therefore says little about whether the next speaker will be transcribed well. Severity of the disease, microphone distance and speaking rate probably matter, but this data does not record them, so we cannot say which.

What this data cannot tell you

• One language and one region. Every recording is Italian, from one region of Italy. Other languages may behave differently. For how accent and language change accuracy, see our data on accent robustness.
• Read text only. Speakers read a fixed passage and fixed sentences. Dictating your own words, with pauses and corrections, is a different task, and the set's unscripted speech is not part of this analysis.
• Disease stage. Speakers with more advanced Parkinson's may fare differently from the people recorded here.
• Small groups. With 15 to 24 speakers per group, the intervals are wide. A difference that includes zero is not a proof of no difference.
• Three engines on one Mac. We tested on-device engines only. For speed and accuracy of on-device engines in general, see the on-device benchmark. Cloud engines are not covered.
• Repeat sessions. Some people were recorded twice, so the sessions are not fully independent.

What this means if you have Parkinson's and want to dictate

Expect more errors than a healthy speaker of your age gets, and expect them to be uneven: most of a passage may come out well, while a few stretches come out badly. None of the three engines was measurably better for people with Parkinson's, so there is no engine to pick from this data alone. Choose on the other things that matter to you, such as whether audio stays on your device, and test with your own voice.

A few practical steps follow from the data:

1. Try each candidate engine on a paragraph of your own reading, since averages hide individual variation.
2. Dictate in short sentences if long passages go wrong for you. The short-sentence figures were lower for all three engines, though the data cannot prove the length is the cause.
3. Plan to proofread. The worst recordings in the set were far from usable.
4. If errors stay high, voice-trained or adaptive tools exist for speech impairments. Our guide to speech recognition for speech impairments covers them, and our overview of accessibility covers the wider set of options, including typing alternatives.