Wega Labs · The destination
Thought to text is the destination. The engineering path runs through selection, silent speech, and non-invasive neural control. We are building the tractable steps now.
Definitions / 01
Most confusion about this field comes from collapsing three separate technologies into one phrase. They differ by orders of magnitude in bandwidth, in invasiveness, and in how close they are to being a product a person could own.
Surface EMG is the closest consumer path today; its signal is the neuromuscular act of speech.
Signal bands / 02
Separating the three bands by bandwidth, invasiveness, and product readiness makes the engineering path clear.
Sub-vocal articulation produces measurable muscle activity in the jaw, throat, and face even when no sound is made. This is the closest band to a consumer path today. Its signal is the neuromuscular act of speech, which defines the near-term product boundary.
Published work recovers coarse semantic categories and, under fMRI, approximate continuous meaning. Bandwidth is low, sessions are long, and per-subject training is heavy. Its present value is as a research signal; interface-grade bandwidth remains the engineering target.
Implanted arrays in clinical populations produce the highest-bandwidth results in the literature. They define the frontier the field measures against; our programme concentrates on non-invasive systems.
State of the art / 03
Every figure below is reproduced from its citation and checked against it. Results from other teams remain attributed at the row.
Requires a multi-tonne magnetoencephalography machine in a magnetically shielded room — and the participant is physically typing on a QWERTY keyboard. It decodes motor commands to the fingers, not inner speech.
Two of every three characters wrong. This is the number that matters for anything you could actually wear, and it is not close to usable.
Genuinely the closest thing to silent speech. Constrained vocabularies, no announced price, no ship date, and the current build still tethers to external processing.
It works and you can buy it. But it is bundle-locked to one ecosystem with no third-party SDK, so nothing independent can be built on it.
Sources · Meta, Brain-to-Text Decoding: A Non-invasive Approach via Typing (arXiv 2502.17480) · Kapur et al., AlterEgo (IUI 2018) · Meta Quest engineering blog, CES 2026. Figures are reproduced as published and are not measurements of our own systems.
The finding the field under-reports / 04
The artifactual component of the EEG signal is significantly more informative than brain activity with respect to classification accuracy — consistent across different feature extraction methods and classification pipelines.Artifacts in EEG-Based BCI Therapies: Friend or Foe? · Sensors 22(1):96
When researchers separated artifact components from brain-signal components in EEG recordings, the artifacts — jaw, eye and neck muscle activity — were consistently more predictive of intended movement than the brain signal was.
Movement-related artifacts are not random noise. They contaminate the signal of interest in a predictable way, which makes them more useful to an automated classifier, not less. Most published BCI work does not report whether it controlled for this.
The paper draws the distinction that matters: informative artifacts are a helpful friend in communication applications, and a serious foe when the goal is estimating a physiological brain state. A decoder optimised purely for accuracy will happily select for the clench.
We consider this the single most important fact for anyone entering this field, ourselves included. It is why one of our five public commitments is an artifact control.
Prior art / 05
The most concrete public prototype of this idea is ThinkVoice, built at the Grokathon. It deserves credit, and it deserves an accurate description.
Someone speaks. A language model generates candidate replies as on-screen chips. The user selects one by tapping or with a blink, nod, or head shake through an optional head-worn sensing device or the phone’s on-device face camera. A cloned voice says the chosen line aloud.
That is roughly 1.6 bits of human intent per utterance. The model supplies everything else. Its current public page describes invite-only TestFlight software for iPhone and Apple Watch, not a public App Store release, and does not establish neural thought decoding. It demonstrates the selection architecture cleanly.
Plate VI / 06
select → screen_03 · select → dark · confirm
attend → screen_03 · select → dark
Research standard / 07
We name the signal, input method, and model contribution.
We report target count and time per selection beside accuracy.
We test every neural result against a muscular explanation. Mixed signals remain labelled as mixed.
Prior results keep their citation, apparatus, participants, and conditions.
We publish each gate, result, and dataset.
Programme / 08
Selection is the first rung toward thought to text.
Thought to text is the destination and the reason this lab exists. The programme advances through measurable layers of bandwidth, control, and product readiness.
The rung a person can actually use today is selection — a handful of bits, aimed at a model that supplies the rest. That is where our engineering is, and the counter on the front page reports exactly what it costs.
Where the work is now
The interaction earns its place before neural hardware enters the loop. A laptop camera and a live information counter answer that question first.