When voice assistants do not understand you.

Cut off, and misheard

In 2022 Apple researchers surveyed 61 people in the United States who stutter. Of those who knew voice assistants, 61 percent said assistants cut off their speech before they finished, and 58 percent that assistants did not understand them well enough; more than half said their speech was recognised rarely, sometimes or not at all. Some had other reasons to hold back: 37 percent did not want others to hear them talking to a device. When dictating, repeated words and fillers ended up in the text, which people disliked.

What changed when the researchers adjusted the system

The team then recorded 91 people who stutter saying assistant commands and dictation, and tried three changes on a speech recogniser:

• Waiting longer before deciding you are done. The part of the system that decides when you have finished cut off almost a quarter of utterances; tuned for people who stutter, it cut off far fewer, at the cost of a short extra pause before the answer.
• Leaning more on likely sentences, by tuning how the recogniser weighs what it hears against what people usually say.
• Cleaning the transcript, removing fillers such as “um” and repeated words or phrases.

Together, Apple reports, the system cut utterances off 79.1 percent less often and improved the average word error rate from 25.4 to 9.9 percent. Most participants in that test had moderate or severe stutters, and the recordings date from 2020 and 2021.

A shared library of voices

The Speech Accessibility Project, run by the University of Illinois Urbana-Champaign with support from Amazon, Apple, Google, Meta and Microsoft, collects recordings from people whose speech is affected by ALS, Parkinson’s disease, cerebral palsy, Down syndrome, stroke, stuttering or hearing loss. By the end of September 2025 about 2,000 people had taken part. The recordings are de-identified and shared with the sponsoring companies and with researchers who sign a data use agreement, on the condition that they do not try to identify anyone. It currently recruits adults in the United States and Canada who stutter or are deaf or hard of hearing, except in Illinois, Texas and Washington, where state privacy laws stand in the way.

Its first published study trained a recogniser on 151 hours of speech from people with Parkinson’s. On new recordings, it got 23.7 percent of words wrong, against 36.3 percent for a system trained on speech from people without Parkinson’s. Speakers who talked unusually fast or stuttered saw smaller gains.

Settings that can help today

• On a Mac, Siri can listen for atypical speech: Apple menu, System Settings, Accessibility, Siri. Apple says it may help with conditions such as cerebral palsy, ALS or a stroke, and that it is available in U.S. English.
• On Android, Google’s Voice Match teaches the assistant your voice; Google says it helps recognise you better in noisy places. In the Gemini app it sits under Settings, Talk to Gemini hands-free, Manage Voice Match.
• Google’s Project Relate, an app that learned an individual’s non-standard speech from recorded phrases, is a beta that is not accepting new users; existing users can keep using it.
• Look for a longer wait. If your assistant lets it wait longer before it answers, try that first: in Apple’s study, waiting longer was what stopped most interruptions.
• Type when speaking is tiring. Many assistants take typed requests too.

A button you hold

Most assistants have to guess when you have finished. A hold-to-talk button removes the guess: the device listens for as long as you hold it and stops when you let go, however long your pauses are. That is how Lychnos Ring is designed: hold to talk versus wake words. Understanding every voice, with its stutters, lisps, accents and mumbles, is the first thing we are working on for the ring, ahead of everything else. It is in development, and we will say what works only once we have measured it with the people it is for. Accents have their own story: voice assistants and accents.

Sources, read 5 November 2026: Apple Machine Learning Research, from user perceptions to technical improvement: enabling people who stutter to better use speech recognition (Lea and colleagues, CHI 2023) and improved speech recognition for people who stutter (May 2023); Speech Accessibility Project, its about and FAQ pages; University of Illinois, automatic speech recognition learned to understand people with Parkinson’s disease, 21 October 2024; Apple, change Siri settings for accessibility on Mac; Google, set up and manage Voice Match and Project Relate.