Is speech recognition less accurate for older adults?

Yes. In our October 2026 test, every engine made more errors on healthy speakers aged 60 to 77 than on healthy speakers aged 19 to 29 reading the same passage, and for all three the difference is beyond chance. Word error rate is the share of words that are wrong, missing or added, so lower is better.

Parakeet TDT v3 went from 1.79 percent ¹ for the young group to 8.00 percent ² for the older group. Whisper large-v3-turbo went from 3.04 percent ³ to 6.84 percent ⁴. Apple SpeechTranscriber went from 3.89 percent ⁵ to 8.82 percent ⁶.

The older group had higher error rates on all three engines

The comparison uses the Italian Parkinson's Voice and Speech set. Its two healthy groups are 15 speakers aged 19 to 29 and 21 speakers aged 60 to 77. Each person read a short passage twice, and we ran the recordings through the three engines on a Mac with an M2 Pro.

The chart also shows the people with Parkinson's disease. For this question, compare the first two groups. The rows below show the same figures as a table.

The gap is beyond chance, but its size is uncertain

For each engine we calculated the difference between the older and the young group and a 95% interval around it, widened to allow for the 9 comparisons the test makes in total. If the interval includes zero, the test cannot tell the difference from chance. For all three engines the interval excludes zero, so the extra errors are a measured difference and not luck of the draw.

The intervals are wide, though. The lower end of the range for Whisper and for Apple is only modestly above zero, so the real gap could be small for them. In points, Parakeet's gap is the largest of the three ⁷ and Whisper's is the smallest ⁸. We did not test whether the gaps differ from one engine to another.

Engines are mostly level for older voices

Among the older speakers, Whisper had the lowest error rate (6.84 percent), Parakeet was next (8.00 percent) and Apple was highest (8.82 percent). Only one of those differences is measured: Whisper made fewer errors than Apple ⁹. Whisper and Parakeet are level ¹⁰, and so are Parakeet and Apple ¹¹, so the rest of the order among the older speakers is not a result. Among the young speakers, Parakeet made fewer errors than Apple ¹² and than Whisper ¹³, while Whisper and Apple are level ¹⁴. Parakeet had the lowest error rate for young voices and the largest gap in points when voices were older, though the gaps themselves were not compared between engines.

Individuals vary more than averages suggest. On every engine, the worst older speaker had an error rate well above the group average, and the older group's worst readings are well above the worst young readings. Age alone does not tell you how a given person will fare.

The older speakers also read ten short sentences

The older healthy speakers also read ten short sentences. Whisper had 4.04 percent ¹⁵, Apple 5.30 percent ¹⁶ and Parakeet 5.30 percent ¹⁷. The young speakers did not read these sentences, so there is no age comparison for them. On these sentences the three engines are level with one another ^{18 19 20}. Each engine scored lower on the sentences than on the passage for the same older speakers, but the texts differ, so that is not evidence that short sentences are easier.

What the numbers do and do not measure

• This is a comparison of two groups of different people, not the same people followed as they age.
• The oldest speaker in the older group is 77, so the results say nothing about people in their eighties or nineties.
• Everyone read a fixed text in Italian, from one region. Spontaneous speech, other languages and accents may behave differently.
• We do not know how much of the gap comes from the voices and how much from recording conditions, hearing, reading fluency or other factors, since the test did not separate them.
• The engines ran on a Mac. Cloud services were not tested.

What it means if you dictate as an older adult

1. Expect a few more corrections. The gap is real in our test, though a higher error rate is not the same as unusable text. Read the text back before you send it.
2. Try more than one engine on your own voice. On the passage, Parakeet made fewer errors than the other two for young voices, but for older voices only Whisper against Apple was a measured difference. Our comparison of dictation products lists the tools side by side.
3. Do not assume age is the cause of a bad result. Microphone placement, background noise and speaking pace matter too, though our data does not test them.

For a related question, see whether dictation works for people with Parkinson's disease, where healthy older speakers are the comparison group. The full results are in the full dataset. More on speech that engines find hard is in the accessibility overview and our guide to speech recognition for speech impairments. Related data: accent robustness and an on-device speech recognition benchmark.