What is silent speech? How it works, and what you can actually buy

Four things that get called "silent speech"

The term covers several different kinds of input. Knowing which one a product means tells you a lot about how it works and how close it is to market.

Whispering still produces sound, so it is not truly silent. Imagined speech requires surgery today. Most consumer silent-speech products aim at the first and third types.

How silent speech works: the main sensing methods

Every silent-speech system has two parts: a sensor that captures something about how you form words, and a model that maps those signals to words. The sensor sets the limits. Here are the approaches in use, with a real example of each.

The full, dated list of products and projects is on our tracker.

Why lip reading alone is hard

Many sounds look the same on the lips. "Pat," "bat" and "mat" are made with the same mouth shape, and much of speech happens out of sight, at the tongue and soft palate. Researchers describe speech as having roughly 40 distinct sounds (phonemes) but only about 12 to 15 distinct visible mouth shapes (visemes). That is why camera-only systems depend heavily on language models to guess from context, and why newer approaches try to sense the tongue and inner mouth with radio, sonar or ultrasound.

Who it is for

• People who have lost their voice. After a laryngectomy, a stroke that causes dysarthria, or with ALS, silent-speech devices could restore communication. Much of the strongest published evidence, such as Cambridge's Revoice and the brain-implant studies, comes from this medical work.
• Privacy in public. Dictating or talking to an AI assistant on a train or in an open office without being overheard.
• Noisy places. Factories, flight lines or crowds where microphones struggle.
• Hands-free control. Short commands to a computer, phone or headset without speaking aloud.

Can you buy a silent-speech device today?

Not a finished one, as of November 9, 2026. The options look like this:

• Preorder: Subvocal's glasses take a $199 refundable deposit toward a $499 early-bird price, for US addresses only, with shipping estimated for Q1 2027 (Subvocal preorder terms).
• Waitlist: AlterEgo, the MIT Media Lab spin-out, shows a pre-production prototype and a sign-up form, with no price (AlterEgo).
• Whisper-level, not silent: Subtle's Voicebuds earbuds are built to pick up speech "below a whisper," but they still rely on sound (Subtle).
• Big tech: Apple bought Q.ai, a company with a silent-speech patent application, but has not announced any silent-speech feature.

Our stage labels explain the difference between research, prototype, preorder and shipping.

How to read accuracy claims

Silent-speech accuracy numbers are easy to misread. Before trusting one, check four things:

• Vocabulary size. 95% accuracy on 30 fixed commands is a very different achievement from 95% on open dictation.
• Which users. Results on the people a model was trained on are usually much better than results on new users.
• Calibration. How many sentences or minutes of training each new user needs.
• Who measured it. A peer-reviewed study or an independent test counts for more than a company demo video.

Word error rate (WER) is the standard measure for dictation: the share of words the system gets wrong. Lower is better.

Is silent speech private?

Not automatically. Silent speech keeps your words from people nearby, but the signals still have to be decoded somewhere. If that happens on a company's servers, privacy depends on how the company handles the data. Subvocal, for example, has said decoding runs in the cloud and that its product privacy notice will come later. Read our report.

Silent-speech devices also do not read thoughts. Consumer systems decode words you deliberately articulate with your mouth or throat. Decoding imagined speech has only been shown with surgically implanted electrodes, in clinical research, and the Stanford team built a mental "password" so that decoding stays locked unless the user chooses to unlock it.