The word you didn't say

Read this sentence and listen to yourself. Most people hear a voice saying the words in their head, and that voice has a body. When you subvocalize, the speech-motor cortex drafts the muscle commands for the words you are not speaking. The mouth mostly stays still, but the plan exists, electrically, milliseconds before the point where execution was cancelled. A word you never said still began life as a signal. This note is about the design pattern that follows from that fact, because it is the pattern our entire product surface sits on.

The spectrum of saying

Saying a word is not one event but a pipeline: the cortex assembles the plan, the motor system executes it, muscles contract, air moves, sound results. Interfaces have always lived at the end of that pipeline, catching the sound or catching the fingers. The spectrum runs:

Overt speech, said aloud. Strongest signal, trivially labelable, socially loud.

Mouthed speech, lips and tongue moving with no voice. The muscles still fire, so surface sensors on the jaw and throat can read words directly from the electrical activity of the muscles themselves.

Subvocal and imagined speech, nothing moves that a camera could catch. The plan is formed and withheld. The signal retreats into the cortex, faint but present over the speech-motor areas, and into microvolt-level residue in the muscles.

The further down the spectrum, the quieter the signal and the more private the act. That trade, signal strength against social invisibility, is the design space.

The pattern, stated plainly

Problem: the human output channel is motor, serial and socially conspicuous. Typing needs hands and eyes; speaking needs a voice and an audience of at least one; both are impossible while driving, holding tools, or standing in a conversation.

Solution: move the tap upstream of execution. Sense the speech plan where it is drafted, decode it into candidate words, and let a confirmation gate, a whisper, a glance, a voice prompt, decide what becomes action.

Why it works: the drafting stage is shared. The cortical plan for a word spoken aloud and the plan for the same word merely imagined overlap heavily, which creates the training bridge that makes the pattern practical. You do not need people to imagine thousands of words into a vacuum with perfect labels. You record them saying thousands of words, where the audio supplies free, perfect labels and the signal is at its strongest, and then you adapt the decoder across the shift to the unspoken version, where the labels are the same words in the same order from the same session.

The pattern is: train on the word that was said, deploy on the word that wasn't.

The evidence trail

This is not a new dream, and the trail shows exactly how far it has been pushed. NASA's Ames group demonstrated it in the early 2000s with surface muscle sensors on the throat, recognizing small command vocabularies subvocally, numbers and rover commands, well enough to control machines. MIT's AlterEgo made it wearable in 2018: a pendant of electromyography sensors plus bone-conduction output let a user ask silent questions and hear silent answers, with around ninety percent accuracy on a closed vocabulary of commands and chess queries. On the cortical side, the invasive studies proved the planning stage decodes: patients who can no longer move at all, attempting to handwrite or attempting to speak, produce plans rich enough to reach sentences. What remains unproven anywhere is open-vocabulary silent speech from a wearable, non-invasive cap. Closed vocabulary is demonstrated. Open vocabulary is the open lane, and it is blocked the same way everything in this field is blocked: by the scarcity of paired data.

Why this pattern fits our moments

Look back at the three product moments and the fit is exact. Driving: eyes on the road, hands on the wheel, voice reserved for passengers; a subvocal thought plus a voice confirmation is the only input that costs nothing. Hands busy, on tools, mid-conversation: the same arithmetic. Even the sleep moment is the pattern in reverse, capture without expression, reading the overnight trace of the day's words. In every case the interface disappears not because it was miniaturized but because it moved to a channel that was never occupied: the draft, not the delivery.

How we will train it

The protocol writes itself from the pattern, and it is our Phase 1 design. Each recording session contains three passes over the same sentence set: spoken aloud, mouthed silently, and purely imagined. The aloud pass anchors labels and calibrates the strongest signal. The mouthed pass bridges toward silence while muscles still corroborate the timing. The imagined pass is the target domain. One decoder, adapted across three rungs of the same ladder, evaluated with the honesty rules from our note on reading decoding papers: cross-session, cross-subject, chance level stated. The confidence gate comes from our agent note: above threshold, act; below, ask. Subvocalization is quiet by nature, so the confirmation must be quiet too.

The whole plan of this company, compressed into one sentence: the input layer for machines has always listened at the mouth and the hands, and we are moving it up to the place where the words are still being formed. The word you didn't say is the one we are learning to hear.