Signal
Silent speech is leaving the lab: words formed without a sound are being turned into text and commands. AlterEgo, a startup out of MIT, unveiled a wearable in September 2025. It reads the faint muscle signals of words you form silently in the jaw and throat, then answers by bone conduction, a vibration carried through the skull to the inner ear, so that only you can hear it. A year on, its site still shows a pre-production prototype and a sign-up form. But the labs are closing in on everyday accuracy. A team from Cambridge and Beihang University in Beijing put fabric sensors on the throat and a language model behind them, and stroke patients who could no longer speak clearly got their sentences back with four errors in every hundred words. In Japan, a soft patch worn on the hand, held to the lips only when needed, recognised a thirty-word vocabulary with 97 percent accuracy and even flew a drone. Engineers in Taiwan reached an error rate of 8.5 percent on a vocabulary of 1,396 words with microneedle electrodes on the face. The same reading of muscle signals is already on sale at the wrist: Meta's Neural Band, sold with its display glasses, turns finger-writing on any surface into messages, at about 21 words a minute in the company's own study. Typing taught people to talk to machines with their fingers. The next way goes straight from the throat, and nobody in the room will notice.
So what?
Silent speech is becoming an input like touch. The fastest way to tell a machine something is the sentence you never say out loud.
Evidence
Decrypt, 8 September 2025 (launch): Boston startup AlterEgo unveiled a wearable that lets people communicate silently with machines by decoding muscle signals in the jaw and throat as they form words without speaking, with responses delivered privately through bone-conduction audio. It builds on a 2018 MIT Media Lab prototype, and the company is yet to give a launch date, a price or its funding.
Nature Communications, 19 January 2026 (study): a team from the University of Cambridge and Beihang University in Beijing built a throat patch of fabric sensors that picks up the vibration of the voice-box muscles and the pulse in the neck, and a large language model assembles and corrects the sentences from what the patch decodes. Five stroke patients whose speech had become slurred reached a 4.2 percent word error rate on 47 Chinese words and 20 sentences.
Advanced Intelligent Systems, 9 June 2026 (study): Yokohama National University and GMO Pepabo's research institute built a soft muscle-signal sensor worn on the hand, with a fingertip electrode you bring to your lips only when you need it. A neural network recognised a 30-word silent vocabulary with 97.2 percent accuracy across three people, and the team steered a drone with silent commands, the company said on 23 June 2026.
Worth asking
If you could write as fast as you think, with nobody around you hearing a word, where in your week would you do it first?
