Which speech recognition engine understands Parkinson's speech best?

No engine was measurably better on either task. On a read passage Apple SpeechTranscriber had the lowest error rate for people with Parkinson's disease, and on ten short sentences Whisper large-v3-turbo did, but the three engines are level in both. On the passage, Apple had a word error rate of 13.81 percent ¹, Parakeet TDT v3 had 14.32 percent ², and Whisper had 21.84 percent ³. Word error rate is the share of words that are wrong, missing or added, so lower is better.

Within the Parkinson's group the test measures no difference between engines on either task. Whisper's passage rate is several points higher than the other two, but that gap is not a measured difference: its interval, allowing for the engine comparisons, includes zero because its errors sit in a few recordings ( ⁴, ⁵).

On the passage, Whisper's error rate was highest but the engines were level

Apple and Parakeet were close, and Whisper's rate was several points higher. The test covers 24 people with Parkinson's disease aged 40 to 80 reading a short passage, 52 readings in all, on a Mac with an M2 Pro.

The test compared the engines with each other on the same recordings. For people with Parkinson's disease none of the differences is measured, so treat all three as level, Apple against Parakeet ⁶ as much as Whisper against either. The engine differences that are measured sit in the healthy groups on this passage: for young speakers Parakeet made fewer errors than Apple ⁷ and than Whisper ⁸, and for healthy speakers aged 60 to 77 Whisper made fewer than Apple ⁹. We can also say how each engine fares against its own result for healthy speakers aged 60 to 77 on the same passage:

The intervals are 95% intervals that allow for the test's 9 comparisons. Whisper's gap to healthy speakers is the widest in points: it had the lowest rate of the three for healthy older speakers, measurably below Apple's, and the highest for people with Parkinson's, where the engines are level. Apple's gap is the smallest in points, but the test cannot tell it apart from chance, so we would not claim Apple handles Parkinson's better than it handles healthy older voices. The honest reading is that Apple's result is a good one, with wide uncertainty.

On short sentences the three engines were within a narrow band

When the same people read ten short sentences, the three engines were close: Whisper 7.47 percent ¹⁰, Parakeet 7.62 percent ¹¹ and Apple 8.27 percent ¹². Whisper's rate is the lowest in points, but no engine difference is measured here ( ¹³, ¹⁴), so it does not carry much weight.

For these sentences, every engine's gap to healthy speakers aged 60 to 77 has an interval that includes zero. Parakeet's gap was the smallest in points and Whisper's the largest, but the test cannot tell any of the three from chance. For comparison, healthy older speakers scored 4.04 percent ¹⁵ (Whisper), 5.30 percent ¹⁶ (Apple) and 5.30 percent ¹⁷ (Parakeet).

Averages hide large differences between people

The average is pulled up by a minority of speakers. On the passage, the worst single speaker had an error rate several times the group average on every engine. An engine that does well on average can still fail badly for you, so the table is a starting point and not a recommendation.

What the numbers do and do not measure

• People read fixed text. Your own dictation, with its own words, pauses and corrections, may behave differently.
• All speakers are Italian, from one region, and the engines were run in Italian. Another language may rank the engines differently.
• The test has no measure of disease severity, and more advanced Parkinson's may fare differently.
• Only on-device engines running on a Mac were tested, not cloud services. Engine versions and settings change, so the results are a snapshot from October 2026.
• With 22 to 24 people per group, small gaps between engines are within the noise.

What it means for choosing a tool

1. If you dictate longer text, start with Apple's built-in engine or Parakeet. Both had lower error rates than Whisper on the passage in our data, though the test does not measure that as a difference between engines.
2. If you dictate short phrases, any of the three is a reasonable start. The differences were small and none is measured.
3. Test on your own voice. Record a paragraph of your own and compare two engines. Our comparison of dictation products puts the tools side by side.
4. Always read the text back. Even the lowest average for this group is well above what healthy older speakers scored on the passage.

See whether dictation works for people with Parkinson's disease for the comparison with healthy speakers, and the question on older adults for the effect of age alone. The figures are all in the full dataset. For wider context, read the accessibility overview and the guide to speech recognition for speech impairments. Related data: accent robustness and an on-device speech recognition benchmark.