Why do voice assistants misunderstand you?

Voice assistants misunderstand you for three main reasons. The recognizer may hear the wrong words, because the sound is noisy, distant or unlike the speech it handles best. The assistant may read the words correctly but pick the wrong meaning. Or it may decide you have finished before you have, and act on part of what you said. Each cause leaves a different trace, so a few minutes of checking will tell you which one you are dealing with.

Background noise and distance corrupt the audio first

An assistant can only work with what reaches its microphone. A 2017 paper on distant speech recognition says the technology still lacks robustness when non-stationary noise and acoustic reverberation are present ¹. A TV, a fan, a second voice or the echo of a bare room all add to the signal the recognizer has to separate from your speech.

Apple's advice for Mac Dictation is a useful checklist for any voice input. Apple tells you to position yourself as you normally use your Mac, to speak clearly at a volume that is neither too quiet nor too loud, and to avoid background noise ². If you follow that advice and the words still come out wrong, the problem is probably not the software's manners. It is the signal.

The symptom is recognizable: the assistant returns words that sound nothing like what you said. If those errors happen when the TV is on or a door is open, test it. Move closer, lower the background sound, and repeat the same request. If the result improves, the audio was the cause.

Accents and unusual pronunciation fall outside what the model handles best

Recognition is least reliable for speech that is less like the training data. In the EdAcc corpus of accented English, the best English model in a 2023 study averaged a word error rate of 19.7 percent on conversations between friends ³. For the same model, the conversational rate was about 7.3 times its rate on US English clean read speech ⁴. Conversation is looser than reading aloud, so the test mixes accent with speaking style. It still shows that accuracy can drop sharply once speech moves away from the most common patterns.

The gap also appears between speaker groups. In a 2020 study of five commercial speech recognition systems on interview recordings, the average word error rate was 0.35 for Black speakers and 0.19 for white speakers ⁵. The study is about groups, not individuals, but it means a failure can come from the system rather than from how you spoke.

Unusual pronunciation is a related problem. A review of pronunciation teaching says that systems such as those on Siri or Alexa treat a learner's mispronunciations as noise, because their goal is to transcribe the sounds the speaker intended ⁶. In practice, a name, a brand or a technical term said in a way the system does not expect can come out as a more common word. For speech that is harder to produce, see the question on whether voice recognition can understand dysarthria.

Assistants cut you off and then act on part of what you said

Mishearing is only one failure. The assistant also has to decide when your request is complete. In a 2023 study by Colin Lea and colleagues, 61 percent of people who stutter and were familiar with voice assistants named being cut off before they finished talking as a reason they used them less often ⁷. That figure describes that group, not every speaker. A pause in the middle of a sentence can still look like the end of a request, so the assistant may act on the first half.

A correct transcript can still produce the wrong action. The assistant maps the words to an intent, such as a timer, a call or a search. If the request could mean two things, the assistant may pick the one you did not want, even though it heard every word.

Which cause fits which symptom

What to do when an assistant misunderstands you

1. Check the words first. Look at the transcript or the text the assistant shows before you judge the action. If the words are wrong, the cause is the audio or the pronunciation. If the words are right, the cause is the meaning or the cut-off.
2. Fix the sound before you repeat the request. Stand where you normally use the device, reduce the TV or fan, and speak at a normal volume.
3. Give a complete, specific request. A request with the app, the person or the object named is harder to misread than a fragment.
4. Pause only when you mean to stop. A short pause mid-sentence can end the request early. For a long message, keep speaking and check the result afterwards.
5. Use dictation for text you will keep. Assistants are built to answer requests, not to take long text. The guide on assistants and dictation sets out which tool fits which job, and the question on whether assistants can take long dictation covers the specifics. If you dictate on a Mac or Windows PC, Yap, which we make, is an on-device dictation app for both.

How far the evidence goes

Read the same sentence twice in the same room, once at normal volume and once with the TV on, and count the wrong words each time. If the errors rise with the noise, the fix is in your setup. If they stay the same, the cause is more likely your accent or the recognizer itself.

The studies above used particular recordings, surveys and groups of speakers, and vendors change their recognizers over time. We did not find a current independent test that compares the error rates of Siri, Alexa and Gemini on the same speech. The glossary entry on word error rate explains how that rate is measured. For a device that wakes without you, see the question on how often voice assistants activate by accident, and for accents specifically, see the question on why dictation does not understand your accent.