Convergence Watch: Four Sensors Are Racing to Read the Words You Don’t Say Out Loud

On January 29, 2026, Apple paid roughly $1.6 billion for seventeen US patents and seventeen employees. The patents do not describe a chip, a screen, or a battery. They describe shining a laser at the side of your face and reading the words you are about to say — before you say them.

The startup was called Q.ai, formerly Q (Cue) Ltd., founded in Tel Aviv in 2022 by Aviad Maizels, the same engineer who sold Apple PrimeSense in 2013 for around $350 million. PrimeSense built the 3D structured-light camera that later became Face ID. Q.ai built a tiny coherent light source that bounces off your cheek, catches the speckle pattern returning from skin pulled microns by underlying muscle, and decodes that into language. Apple has now bought, twice, from the same founder, the input layer of two consecutive computing eras. The deal is Apple’s second-largest acquisition in its history, behind only Beats.

That is the hook. The story is bigger.

The bidding war for the words you don’t say

Filter the US grant feed for any patent whose abstract mentions silent speech, subvocalization, inner speech, EMG-coupled speech, or facial skin micromovements. From 1998 through 2023, the patent office issued twelve such grants. From January 2024 through May 12, 2026, it issued thirty-one. The category has tripled its entire historical output in twenty-eight months.

What is striking is who is filing. Q.ai accounts for fourteen of those thirty-one, which is almost all of its US grant portfolio. Snap Inc. owns four EMG-speech grants, plus a broader stack of speech-related filings around its AR eyewear platform. Wispr AI, currently in talks at a $2 billion valuation per Bloomberg’s May 12 report, has two grants. Meta’s Reality Labs surfaces through older CTRL-Labs filings on wrist-based neuromuscular sensing, the gestural cousin of the same problem. Ericsson holds a 2024 grant on subvocal authentication, where you “say” your password without moving your lips and the device reads it off your throat.

Mentally delete the phrase “silent speech” from those abstracts and the patents still belong on the same shelf. They are all measuring articulatory motor signals — the small electrical, mechanical, or skin-deformation traces of words that the brain has sent to the face but the lungs have not yet voiced. This is the coherence test you want a convergence story to pass. It passes.

What does not pass is the assumption that everyone is using the same sensor. They are not. Four very different physical modalities are converging on the same neurological event from four directions.

Mechanism one: laser speckle on the cheek

Read US grant 11922946, assigned to Q (Cue). A wearable directs coherent light — usually infrared — at a small patch of facial skin. A photodiode catches the reflection. As muscle fibers under the skin contract a fraction of a millimeter to form the silent shape of a vowel, the speckle pattern returning to the detector shifts in a way a neural network can learn. The system reconstructs phonemes from skin motion smaller than the eye can see. Q.ai filed variants for earbuds (US 12216750), continuous biometric authentication (US 12131739), even noise suppression where the wristband’s optical channel tells a microphone which audio is yours and which is somebody else’s. The technology generalizes because skin is everywhere. Earbuds, glasses, neckbands — anywhere a small light source can stare at a face.

This is what Apple bought. Maizels and his cofounders, Yonatan Wexler and Avi Barliya, are joining Apple along with their team. Calcalist and Globes both reported Apple’s interest was driven by the AirPods and Vision Pro roadmaps, where audible Siri commands have a known social problem: nobody wants to talk to their face computer on the subway.

Mechanism two: muscle electricity on the wrist

Meta’s Connect 2025, on September 17, was the day the Meta Neural Band shipped at $799 with the Ray-Ban Display glasses. The wristband does not look at your face. It picks up the surface electromyography signal at your wrist — the same nerve impulses that drive your fingers to type, except the fingers do not have to move enough for anyone watching to see. In a Nature paper published July 23, 2025 (A generic non-invasive neuromotor interface for human-computer interaction), Meta’s Reality Labs reported handwriting input at 20.9 words per minute, with a further 16 percent gain from per-user personalization.

The CTRL-Labs acquisition was 2019, reportedly for between $500 million and $1 billion. Six years later Meta is shipping it on retail shelves alongside an AR display. Apple, meanwhile, just paid roughly twice what CTRL-Labs cost — for a company with no revenue, no shipped product, and no Nature paper. Apple is paying for a different sensor with arguably better social acceptance: pointing a light at your own face is less weird than wearing a science bracelet.

Mechanism three: facial EMG inside glasses

Snap takes a third path. Read US grant 12118148: subthreshold muscle activation signals from speech-production muscles, picked up by electrodes integrated into a wearable, then mapped by an ML model to inferred speech. US 12525240 patents per-user calibration of that model. US 12517584 patents a transform that removes eye-blink interference from the EMG channel, a tell that Snap is building this for glasses, where blink-noise is loudest. US 12586568, granted March 24, 2026, patents synthetically generating inner-speech training data: turn ordinary spoken-speech recordings into simulated EMG traces, then train the silent-speech decoder on the synthetic corpus.

That is what a company does when it has run out of real subvocal data and needs to feed a production-scale model. Snap’s next, lighter Specs are scheduled for 2026.

Mechanism four: everyone else

The literature carries a fourth and fifth. Cornell’s EchoSpeech (2023, ACM CHI) reads silent speech off acoustic reflections from tiny speakers and microphones built into ordinary-looking eyewear; no electrodes, no laser, just sonar bounced off your jaw. A 2022 Nature Communications paper from KAIST glued ultrathin silicon strain gauges to the throat. Across the OpenAlex corpus, papers indexed against “silent speech interface” rose from 16 in 2019 to 66 in 2025. The academic flywheel is real.

Why now, and who pays

For thirty years the silent-speech literature lived in the assistive-tech aisle: laryngectomy patients, ALS patients, anyone whose larynx had been silenced by disease. That is still a real market, and OpenAlex’s most-cited paper in the field is a 2020 review focused on speech restoration. What changed is that voice assistants got good — and then large language models got good — and the bottleneck moved. You can now talk to a competent agent. You just cannot do it on a bus.

The cost of audible voice input in 2026 is social. The cost of typing on glasses is mechanical: no surface. Silent speech is the only input modality that solves both. That is why a company with no revenue is worth $1.6 billion to Apple, why Meta put a wristband on retail shelves with its Ray-Bans, why Snap is filing inner-speech patents two weeks before its 2026 Specs reveal, and why Wispr — which started as a silent-speech hardware company before pivoting to dictation software — is back in talks at $2 billion despite shipping no sensor of its own.

For an R&D director, the strategic read is narrower than the headlines suggest. Three different sensor stacks are now in serious commercial development for the same input event. The first one to clear an honest 30 words-per-minute in public, without per-user retraining, defines the form factor of the next decade of personal computing. Apple bought optical. Snap is patenting electrical on the face. Meta is shipping electrical on the wrist as the adjacent gesture-input play. The startups that survive will be the ones that pick the surviving sensor before anyone else knows which one wins.

The one nobody is reporting on is the loser’s consolation prize: continuous biometric authentication. Q.ai’s grant on the topic (US 12131739) describes recognizing an individual by the unique pattern of their facial micromovements during ordinary speech. The same hardware that hears your unsaid Siri command can prove, every millisecond, that it is still you wearing the glasses. The bank that licenses that signal first will own a category that does not exist yet.

Method note

Patent counts are drawn from the USPTO weekly bulk grant feed, indexed locally as of May 12, 2026. The “silent-speech” filter is a phrase match in patent abstracts on any of: silent speech, subvocal, subvocalization, inner speech, prevocalization, facial skin micromovement, or the combination of EMG and speech. Assignee totals combine variant company spellings (Q.ai and Q (Cue) Ltd. are treated as one entity) and do not include licensees or co-assignees, so portfolios may be undercounted. Literature counts are from OpenAlex full-text indexing across 357 million scientific works. Apple acquisition pricing reflects the range reported across CNBC, MacRumors, Bloomberg, Cult of Mac, and Calcalist coverage from January 29 to February 4, 2026. Meta Neural Band performance figures are from Meta Reality Labs’s July 2025 Nature paper. Wispr valuation is from Bloomberg, May 12, 2026.