Does speech recognition work if you stutter?

Yes, speech recognition works if you stutter, though it makes errors. Most of the errors are extra words, and many of them appear where you repeat a sound. Sentences can also be cut short when a block or pause looks like the end of what you said. The errors are real, but research has shown they can be reduced. In a 2023 study, technical changes to a consumer recognizer cut its word error rate on people who stutter from 25.4 percent to 9.9 percent ¹.

Most recognizers are trained on typical speech

About 3 million Americans stutter ². Google, which studies these systems, says recognizers are most often trained on typical speech, so people with speech impairments do not get the same benefit ³. Google points to that training data as the cause of the gap.

Most errors are extra words, and many follow repeated sounds

The recognizer's errors are mostly extra words rather than dropped words. In the 2023 study of people who stutter, extra inserted words made up 80.9 percent of the recognizer's errors, while missed words were only 1.6 percent ⁴. The authors found a strong correlation between those insertions and part-word repetitions ⁵. A repeated sound at the start of a word can therefore appear in the transcript as an extra word. When you check a transcript, look for doubled words first.

Blocks and pauses can cut a sentence off before you finish

A recognizer uses an endpointer to decide whether you have finished ⁶, and a block can sound like a pause. The study's baseline endpointer cut off an average of 23.8 percent of the utterances from people who stutter ⁷.

Dictation behaves differently. The authors note that many dictation systems do not use an automatic endpointer ⁸. In systems without one, a pause does not end the input by itself. The study describes dictation as a case where the user finishes by clicking a button again ⁹. For long text, dictation is usually the safer choice than a voice assistant command.

The 9.9 percent result is a research prototype, not a phone app

The 25.4 to 9.9 percent result comes from technical changes that the researchers made to a consumer recognizer, tested with 91 people who stutter ¹. It is a research prototype, not a test of a phone app. We found no published test of Apple's Listen for Atypical Speech or Google's Project Relate on stuttered speech, so no current app has shown this result. Treat the study as evidence that the errors can be cut, not as a promise about any app on your phone.

It depends on the device and on what you say

The result changes with three things. The first is the kind of input: dictated text depends less on automatic endpointing than a voice command does, as above. The second is the device. Where your phone offers it, push-to-talk lets you decide when your utterance ends, so a pause does not end it ¹⁰. Some phones do not offer it, so check your settings. The third is the speaker. We did not find a published figure broken down by how severe a person's stutter is, so your own results may differ from the averages.

What to do

1. Turn on push-to-talk if your phone offers it, so a pause does not end the utterance.
2. Dictate longer text and keep voice commands short. Dictation is less exposed to automatic endpointing than a command.
3. Check the transcript for doubled or extra words and correct those first, since they are the most common error.
4. Test with the words you actually use, not the demo phrases, and keep a keyboard or another way to fix text for the cases that still fail.
5. Speak with a speech-language pathologist if you want help with speech tools. A clinician can assess your speech and suggest what suits it.

For the wider picture, read our guide to speech recognition for stuttering and dysarthria, which covers the research and each app's limits, and our overview of accessibility and health. If your speech is affected by a condition other than stuttering, see whether voice recognition can understand dysarthric speech. For the apps themselves, see which speech recognition apps work for atypical speech. The error rate used in every study here is explained in our word error rate entry, and push-to-talk is defined in the glossary.