The band
A thin sensor band you wear on the forearm or jawline. Surface EMG picks up the muscle commands behind speech; photoplethysmography and a nine-axis IMU supply context. Everything runs on-device.
Hardware detailThought-to-speech / Hardware and software
Speak without a sound.
Lumina reads the intent behind subvocal speech and turns it into action on any machine you're near. No app to open, no command to phrase, nothing on your screen that gives you away.
The gap we're closing
Every interface since the keyboard has asked your hands to do the work.
Voice assistants came close, then asked you to talk to a room. Lumina takes the signal earlier, at the point where you form words but don't say them, and carries it the rest of the way into software.
Non-intrusive by construction
No implant, no surgery, no gel electrodes. A band you put on and forget, reading muscle and nerve activity from the surface of the skin.
Speech, not gestures
Gesture control runs out of vocabulary at about twenty commands. Language is unbounded, and you already know how to use it.
Any software, unmodified
Lumina drives the same inputs your hands would: keyboard, pointer, system APIs. Nothing needs an integration to work.
We build the band and the runtime that makes it useful.
The runtime
Decoding, intent, and execution. The runtime decides whether you meant a prompt, a keystroke, a page navigation, or a command to the lamp in the corner, and then performs it through the host's own interfaces.
Software detailOn the die
The decoding runs here, not on a server.
Every electrode sample is decoded inside the band. Nothing leaves the device until it is words, which is the difference between a microphone pointed at your day and a tool you wear.
- Decode budget
- A quantised recurrent model sized to run continuously inside a low-single-digit watt envelope.
- Signal path
- Eight differential channels at 1 kHz, filtered and scored on-die before any phoneme is proposed.
- Radio
- Only resolved text crosses the Bluetooth link. Raw myoelectric data has no route off the board.
From muscle to machine in four steps.
- 01 · SENSE Surface EMG reads the command. Subvocalised speech still fires the muscles of speech, faintly. A dry electrode array catches that activity at 1 kHz and streams it on-device.
- 02 · DECODE The model reconstructs text. A personal decoder maps your myoelectric patterns to phonemes, then to words. It calibrates to you in minutes and improves as you use it.
- 03 · RESOLVE Intent is matched to the environment. The runtime knows what is in front of you, which window has focus and which devices are in range, and resolves the sentence into a specific action.
- 04 · ACT The action executes. Keystrokes, pointer movement, assistant prompts, browser navigation, or a radio command to a receiver. The host machine sees ordinary input.
If it has a receiver, it's in range.
All of these go through the same runtime. Software never learns that a new interface arrived.
AI assistants
Prompt any assistant without leaving the window you're in.
The web
Navigate, search, fill, buy. Stated once, performed by the browser.
The desktop
Any application on any OS, driven through the inputs it already accepts.
The physical world
Lights, locks, thermostats, blinds: anything with a radio receiver.
What the design refuses to trade away.
A brain-computer interface that changes how you behave has already failed. Lumina is built so that the day you take it off, nothing about your habits has to change back.
-
Nothing on you
No implant, no adhesive, no shave. If it isn't comfortable enough to sleep in, it isn't finished.
-
Nothing leaving the device
Decoding happens on the band. Raw neural and muscle data never needs to reach a server to be useful.
-
Nothing you have to ask twice
The runtime is present in every window, so the same sentence works wherever you are, in whatever application is listening.
We're building with a small group first.
Developers, accessibility engineers, and people who want their hands back. Tell us what you'd use it for.