Is speech recognition less accurate for some people?

Yes. Measured studies consistently find that recognizers make more errors for some groups of speakers than for others. The largest documented gap by race comes from a 2020 study. The biggest gaps in absolute terms affect people with speech impairments. Children and some accents show gaps too, although the evidence on accent is mixed. The cause is mostly the training data, and the gaps can be reduced.

The largest documented racial gap is in five commercial systems

In a 2020 study, researchers ran five commercial speech recognition systems on recordings of Black and white speakers. The average word error rate was 0.35 for Black speakers and 0.19 for white speakers ¹. Those are averages across five systems, not the error of any one product. The study traces the gap to the acoustic models the systems use, which are the part that maps sound to speech units ².

Two limits matter. The study is from 2020, and vendors have changed their systems since. It also averages across five systems, so it does not tell you how any single product performs today.

Speech impairments show the widest gaps

The gaps are largest for speech that the training data barely contains. In Google's Project Euphonia study of 432 people with disordered speech, personalized models reached a median word error rate of 4.6 percent. General models that were not personalized had a median of 31 percent ³.

Parkinson's disease shows the same pattern. A baseline recognizer had a 3.4 percent error rate on typical speakers and a 36.3 percent error rate on speech from people with Parkinson's. Fine-tuning on data from the Speech Accessibility Project lowered the second figure to 23.7 percent ⁴. The fix helped, but a large gap remained.

Google's explanation is direct. Recognizers are most often trained on typical speech, so people with speech impairments or heavy accents do not get the same benefit ⁵. The guide to speech recognition for speech impairments covers what that means in practice, including personalized models.

Children and accents show mixed evidence

A 2025 study of children's speech found that a child's age and the number of words in the audio had the largest effect on accuracy, followed by background noise and pronunciation ability ⁶. If you dictate or caption children's speech, this is a gap to check on your own recordings.

Accent is less clear. A 2004 study of the MALACH oral-history archive found no evidence that accent or speaker age affected recognition ⁷. The same archive study found that signal-to-noise ratio and speaking rate were the two dominant factors behind the overall error, which points to the audio as a major factor alongside the speaker ⁸. The guide on speech recognition and accents goes through the evidence in more detail.

What the gaps mean for you

The practical question is not whether a group is generally served worse, but whether the tool works for your voice and your recordings. A few steps make that measurable:

1. Run the passage test on your own speech. Read a 200-word passage in your usual voice and room, score the output, and repeat it in two or three settings. The accuracy guide sets out the method.
2. Compare more than one tool. Gaps differ between products, and a result from one system does not transfer to another.
3. Improve the audio before you judge the tool. Noise and speaking rate were leading factors in the MALACH archive, and you control both, so test with a closer microphone and a quieter room before you conclude that a tool does not suit a speaker ⁸.
4. Use custom vocabulary for names and terms you use often. The custom vocabulary term explains how such lists work in common tools.

Where a gap affects access to work, school or services, the rights and support available are a legal matter. This page describes what the studies measured and does not give legal advice.

Related reading

For the broader measured picture, see the guide to speech recognition accuracy, and our overview of speech recognition. The companion questions cover whether speech recognition beats humans, why it gets names and jargon wrong and why Whisper makes up words.