Speech Recognition for Stuttering and Dysarthria

Standard speech recognizers make far more errors on stuttered and dysarthric speech than on typical speech, and the most severely affected speakers get the worst results. Google points to training data as the cause. Most recognizers learn from typical speech, so they have little practice with the blocks, repetitions, slurring and imprecise articulation that stuttering and dysarthria produce. Adapting a recognizer to one person's voice closes much of the gap. Google's two studies that compared a personalized model with the general model both found that the personalized one made fewer errors. Apple ships a feature built on this idea, and Google has tested one, but each has limits on device, language and availability, which we cover below. This guide is part of our overview of accessibility and health.

Recognizers are trained mostly on typical speech, so atypical speech gets far more errors

Google, which builds the large recognizers it studies, gives the cause plainly: ¹. The effect shows up in measured error rates. In one Parkinson's study, a baseline recognizer made word errors on 3.4 percent of the words from typical speakers, while the same kind of model made 36.3 percent word errors on speech from people with Parkinson's disease, and fine-tuning it on Speech Accessibility Project recordings cut that to 23.7 percent ². That is the same kind of recognizer, counted the same way. The gap is not a few stray words.

Word error rate is the measure every study here uses. It counts each substituted, inserted or deleted word against the number of words actually spoken. Our glossary entry on word error rate explains how it is calculated.

Speech changes can be an early sign of the conditions that cause them. Google reports that about 25 percent of people with amyotrophic lateral sclerosis (ALS) have slurred speech as their first symptom ³. For those people, the voice features on their devices may be hard to use, which is the gap this section and the rest of the guide describe.

Dysarthria raises error rates sharply, and severity decides how sharply

Dysarthria is a speech disorder with a physical cause. The American Speech-Language-Hearing Association says it happens when brain or nerve damage changes the way your muscles work ⁴. The condition appears in several neurological illnesses, and the estimated share of patients varies widely by study. The association's clinical summary gives these ranges:

The ranges are wide because the association pools several studies, which can define dysarthria differently, so read them as orders of magnitude rather than exact rates.

Severity matters more than the diagnosis. A 2026 study from the University of Strathclyde tested commercial speech-to-text services on the TORGO dysarthric speech corpus and grouped the speakers by severity. The authors note that automatic speech recognition achieves word error rates below 5 percent on typical speech ⁷. On mild dysarthria, the leading systems reached about 1 to 2 percent word error ⁸. On severe dysarthria, every system exceeded 51 percent word error ⁹. The mild and severe groups each had very few speakers ⁸, and the study used default settings, so it shows a pattern rather than a ranking of products.

Personalization changed the picture for the most severe speech. Google trained personalized models for about 430 speakers who had recorded at least 300 utterances ¹⁰. For severely impaired speech, the median word error rate on short home-automation phrases fell from around 89 percent to 13 percent ¹¹.

The three pairs come from different speakers and different tests, so compare the drop within each pair, not the height of the bars across pairs.

Personalized models cut errors sharply, and a few minutes of speech can do much of the work

The most detailed results come from Google's 2019 study of speech from people with ALS and from accented speakers. Relative to the starting model, the personalized models cut word errors by 62 percent on the ALS speech ¹². In absolute terms, the error rate on read message bank phrases fell to 10 percent for mild dysarthria and 20 percent for more serious dysarthria ¹³.

A short training session can do much of the work. The authors report that 71 percent of the improvement on ALS speech came from five minutes of training speech ¹⁴. The paper's results section gives a different figure for the same five minutes, measured against a different reference point, so quote the number with its method rather than as a single figure ¹⁴.

Stuttering responded to technical changes to the recognizer. In a 2023 Apple study of 91 people who stutter, who recorded voice assistant commands and dictation, the authors changed a consumer recognizer and how it handles pauses. The result cut utterances off 79.1 percent less often and reduced word error from 25.4 percent to 9.9 percent ¹⁵. The work included the endpointer, which decides when you have finished speaking ¹⁶.

Personalization has a cost, which is the speaker's own recordings. Google's 2021 study trained a personal model for each speaker who had recorded at least 300 utterances ¹⁰. Shared recordings can lower that cost, and the next paragraph describes one. The 2019 study shows how little personal audio a model needs when it starts from a recognizer already trained on general speech ¹⁴.

Data is what makes personalization possible. The University of Illinois Speech Accessibility Project says its first published study trained a recognizer on 151 hours of recordings from people with Parkinson's disease ¹⁷. The project makes its recordings available to researchers, nonprofits and companies, as the project's article on the Parkinson's study describes. Google's Project Relate, an Android app, was built on research that used more than a million speech samples from participants ¹⁸.

What the research has not measured yet

We could not find a published test of Apple's Listen for Atypical Speech or Google's Project Relate on stuttered speech, so the stuttering results above come from research systems and the consumer recognizer studied in 2023, not from the features you can turn on today. Nor did we find a head-to-head test of these features against the cloud services on the same dysarthric speakers, so the comparison in this guide is between studies, not between products. Each of these gaps is a reason to test a feature on your own speech before you depend on it, using the steps in the last section.

How to read these numbers

Several things limit how far these figures carry. The studies used read phrases from restricted vocabularies, so they measure the training domain, not free conversation or long dictation. Several use very few speakers in each severity group. Most results are for one model at one point in time, and cloud services change. The personalized figures are also often reported as relative improvements, which can sound larger than the absolute change. When you compare results, check whether the figure is absolute or relative, and what the test phrases were.

Stuttering trips recognition at blocks and pauses, and mostly adds words

About 3 million Americans stutter ¹⁹. Stuttering causes recognition errors in two ways, and the first is extra words rather than missing ones. In the Lea study's second phase, which covered 41 participants with moderate or severe dysfluency, the recognizer's errors were overwhelmingly extra words, at 80.9 percent of all errors, against 1.6 percent missed words ²⁰. The authors also found a strong link between those insertions and part-word repetitions ²¹. A repeated sound at the start of a word can come out as an extra word in the transcript.

The second way is cut-off speech. A recognizer decides when you have finished by detecting a pause, and a block or long pause can look like the end of a sentence. The study's baseline endpointer cut off an average of 23.8 percent of the utterances from the participants ¹⁶. In the survey that accompanied the study, 61 percent of the 59 respondents who were familiar with voice assistants named being cut off before they finished talking as a reason they used them less often ²².

Two practical changes follow from this. The first is to avoid voice assistants when you can. Many dictation systems do not use automatic endpointing at all ²³, so dictation in a document or message is less likely to cut you off than a spoken command. The second is to use push-to-talk where your phone offers it. Some smartphones let you hold a button while you speak and release it when you finish ²⁴. With push-to-talk, you decide when the utterance ends, so a pause does not end it.

Apps and settings that help today, and what each one cannot do

The options differ by platform and by how much personalization they offer. This table lists the features we could check today and their limits.

For Listen for Atypical Speech on a Mac, open System Settings, choose Accessibility, then Siri ²⁶. We could not open Apple's iPhone instructions for the same option, so check the matching area on your iPhone or iPad before you rely on it.

Before you rely on any of these features, check four things. Confirm the language and region, since the Mac option is limited to U.S. English. Check whether the feature learns from your voice on the device or sends recordings to a server, which changes what leaves your device. Test it with the phrases you actually use, not the demo phrases. And keep a way to type or use a keyboard for the cases it still misses. Our guide to whether voice to text is safe explains how to check where your audio goes.

Vocal Shortcuts works differently from full dictation. You teach the device a few specific words or sounds and map each one to an action ²⁷, so it suits a small set of commands rather than free dictation.

How to speak to a recognizer, and what to do when it still fails

These steps follow the evidence above. Where a step is only practical advice, we say so.

1. Turn on push-to-talk where your phone offers it, so a pause does not end the utterance ²⁴.
2. Dictate long text and keep commands short. Many dictation systems do not use automatic endpointing ²³. Google's severe-speech result above was measured on short home-automation phrases ¹¹, so short commands are the safer test.
3. Record the phrases a personalized system asks for. Most of the gain in the 2019 study came from a few minutes of speech ¹⁴, so a short training session is worth doing. Project Relate asks testers to record a set of phrases so the app learns their speech ¹⁸.
4. Check the transcript for repeated or extra words. In that study, most errors were insertions, so look for doubled words first ²⁰. This is practical advice rather than a finding about any one product.
5. Use a close microphone. This is general practice rather than a measured result here. Our guide to choosing a dictation microphone covers the options.
6. See a speech-language pathologist if your speech has changed recently. A clinician can assess the speech and suggest tools suited to it. We do not give individual medical advice.

Apple's newsroom describes these tools as designed for people with acquired or progressive conditions that affect speech, such as cerebral palsy, ALS or stroke ²⁵. People who stutter are not named in that list, which is one reason the stuttering evidence above comes from research rather than from the features themselves.

Accents are a separate problem with their own research

Accent is different from dysarthria, and the same method helps both. In the 2019 Google study, personalized models gave a 35 percent relative reduction in word error on accented speech from non-native English speakers ³⁰. The causes of accent errors, and the languages and dialects they affect, need their own treatment. Read our guide to speech recognition and accents, and see the glossary entry on accent for the terms.

Related guides and questions

If you are looking at this topic for a specific need, these questions and guides cover the same ground from a different angle:

• Does speech recognition work if you stutter?
• Can voice recognition understand dysarthric speech?
• Which speech recognition apps work for atypical speech?

Our other guides in the accessibility overview cover related needs: dictation for RSI and carpal tunnel, speech to text for dyslexia and dysgraphia, speech to text for deaf and hard of hearing people, dictation for blind and visually impaired people and dictation for ADHD. For how recognizers are measured in general, see how accurate speech recognition is. For the tradeoffs between on-device and cloud recognition, see on-device and cloud speech recognition.