Subvocal’s silent-speech glasses: how they work and what remains unproven

Talking to an AI is convenient until the people around you can hear the instructions. Subvocal is targeting that gap with glasses designed to turn silent mouth movements into text. Its product FAQ says you deliberately articulate words without producing audible speech; you are still communicating intentionally, just through a different input channel.

The founders describe a prototype that uses radio-frequency sensing to detect facial movement, then an AI model to decode the signal. Their launch post distinguishes the current semi-portable hardware from the smaller, normal-looking glasses they intend to ship. Product images should therefore be read as a design direction, not proof that the finished device is already available.

The company advertises more than 200 words per minute. That is a company-stated speed claim, not a complete measure of usability. The pages reviewed here discuss improving accuracy and correcting mistakes, but do not provide a numerical error-rate benchmark.^[1][2]

The idea, explained

Ordinary dictation starts with sound: you speak, a microphone captures it, and software converts it into text. Subvocal’s founders describe a different route. The glasses send low-power radio signals through the face and measure how those signals change as the muscles used for articulation move. A model maps those patterns to text. The company says the input does not require a camera and does not depend on hearing your voice.

That makes silent speech different from reading arbitrary thoughts. You still have to form the words deliberately. It also differs from camera-based lip-reading: the claimed signal comes from radio sensing rather than video of the mouth. The personal calibration step matters because the system has to accommodate how a particular person articulates.

The founders frame the goal simply: “We wanted the speed of speech without having to make a sound.” But input speed and useful output are different measurements. A fast stream of text may save little time if names, commands, or punctuation require repeated correction. What matters for a real task is how quickly you reach an accurate, usable result, including setup and checking.

There are also two privacy questions. Silent input may address words being overheard in a shared room. It does not, by itself, establish where the resulting text is processed, how long it is retained, or which connected app receives it. Those data-handling questions need their own answers; an acoustic privacy benefit is not a complete privacy guarantee.^[1][2]

“The fastest, most private way to tell them what you want.” Watch at 3:40 ↗

I keep coming back to the input problem. As agents handle more of the execution, explaining what I want can become the bottleneck. Voice helps with that, but saying a detailed work prompt aloud is not always practical in an office or on a train.

That is why the idea interests me. The relevant question is whether silent input can preserve the speed and ease of speaking while working comfortably around other people. A launch demo can show the direction. Reliable accuracy, manageable correction effort, and hardware people can actually wear are what would make it useful day to day.^[1][2]

What to take away

Treat the current offer as a reservation for a product in development. The preorder page lists a $199 deposit toward a $499 early-bird total, with $300 due later. It says deposits are refundable before final payment, shipping is estimated for Q1 2027, and delivery depends on FCC approval. Taxes, shipping, and optional prescription lenses cost extra; a specific delivery date is not confirmed.

To evaluate a device like this, choose a task with a clear result, such as composing a short message containing names and specific instructions. Compare typing, ordinary dictation, and silent input using the same requirements. Count calibration, corrections, and checking alongside generation time. This is a suggested comparison, not a test I am reporting here.

The useful evidence to watch for is a finished hardware demonstration, numerical accuracy results with clear test conditions, correction time, and performance across different users. The 200-plus-words-per-minute claim makes a striking headline. The whole workflow will determine whether it becomes a better way to work with AI.^[1][2][3]