Utter

Type with your
mouth closed.

It feels the tiny muscle signals in your neck and turns them into text. For your messages and your AI agents. No voice, no keyboard.

Join the waitlistSee it in use
Drag to turn it; point attap a part to see what it is. Prototype v1 is the kit you can build today: a soft fabric neckband with 16 dry electrodes inside, 12 on the front of the neck and 2 on each side, plus two beside the printed case at the back, and a lead from each to the case. Concept is the everyday version of the same band, with the leads sewn inside. Head scan: Lee Perry-Smith, CC BY 3.0.
0dB

No sound. Lips closed, nobody hears a word.

8

Silent moves type any word you want.

16ch

Muscle sensors, read 2,000 times a second.

100%

Local. Recognition and the AI run on your computer.

Talking to your phone is loud. Typing pulls your eyes and your hands away. Utter lets you write while you sit in a meeting, ride a train or read your agent's work: you silently move your tongue, and words appear.

A day with Utter.

What you do, what you hear in your earbud, and what appears on screen. Each move takes 1.7 s here, a real pace, but the guesses are luckier than today's: measured, Utter types about 4.5 words a minute in everyday chat and 7.5 replying to a coding agent.

09:40 · In a meeting

Reply to Marta.
Nobody notices.

Your phone stays face down in your lap. You make a few silent moves, and the earbud whispers each word before it's typed. Nine moves, one misheard and undone, then sent.

M
Marta
↑
You, silently–
In your ear–
14:15 · At your desk

Tell your agent what to do.
Keep your hands free.

Your coding agent says a test is failing. You answer without reaching for the keyboard: predictions come from what the agent just said. When it asks permission, one move approves.

agent · ~/code/shop-api
You, silently–
In your ear–
19:05 · On a crowded train

Answer privately.
Without saying it out loud.

Your mum asks if you've landed. The language model already guesses the answer from her question, so it takes four moves, and nobody around you hears a word.

M
Mum
↑
You, silently–
In your ear–
How it works

Eight silent moves.
Any word you want.

1

Wear it

A soft fabric band sits around your neck, with dry sensors on the inside: no gel, on in under a minute. A normal earbud does the talking.

2

Move your tongue, lips closed

No sound, nothing to see. Behind closed lips you press your tongue up, down, left or right, or shape “ka” without saying it. The band reads those muscle signals.

3

Hear the word, and it's typed

The AI on your computer predicts the word you mean and whispers it. One more move accepts it, in any app you can type in.

Your mouth becomes an 8-key keyboard.

Four tongue moves each pick a group of letters, like an old phone keypad: left means “one of n o p q r s”. The AI keeps the words that fit and shows its best guess. Four more moves steer. Most of the time the AI already knows the next word, so you just do ka.

These are the defaults. At setup Utter records 16 moves on you and keeps the 8 it tells apart best: the 7 tongue and jaw moves above, plus 9 silent syllables (ka, tee, loo, yah, ree, thuh, ee, oo, aw), each shaped with the tongue behind closed lips. If a default move reads poorly on you, a syllable takes its place.

Wrong guess? Keep going. Do the group for your next letter and the guess updates. When your word is in the bar, push moves to it and ka types it. jaw undoes the last move.

Pick letters
upa b c d e ftongue tip to the roof of your mouth
downg h i j k l mtongue tip behind your lower teeth
leftn o p q r stongue into your left cheek
rightt u v w x y ztongue into your right cheek
Control
clenchlisten on / offbite your back teeth
pushnext suggestiontongue against your front teeth
jawundodrop your jaw a little
kaaccept, senda silent “k”, back of the tongue

Every move: lips closed, no sound, about half a second.

Example: answer “still coming tonight?” with “yes”. y is in the right group, e in up, s in left. So you do right, up, left: one move per letter, not hunting for the exact letter. “tar” fits those moves too, but after that question the AI guesses “yes”. Then ka sends it. Try it with your keyboard below.

Try it

Your keyboard
is the band.

Each key stands in for one silent move. It's the product's real decoder and word list, running in your browser. Keys run at the band's pace: it reads each move for 0.9 s, answers in your ear, then rests 0.5 s. Go when the voice stops. Too early and it says “too fast”; a key during the rest waits its turn.

  1. Space clench starts listening.
  2. A S D F pick the group of each letter: up abcdef, down ghijklm, left nopqrs, right tuvwxyz.
  3. L ka types the highlighted word. J push moves to the next one. K jaw undoes.
Word guesses come from Utter's own 1.4 MB word model, the same one the app runs offline. It loads with this page and runs on your device.
In your ear–
Built for AI agents

Your agent asks.
You answer, silently.

Agents need short, frequent replies: yes, run it, fix that test. Utter connects to Claude Code, so every reply starts from what the agent just said.

−40%

It knows what was asked

The agent's last message becomes the AI's context. In our benchmark, that cuts the moves per word by about 40%. Replies you sent before come back, even when the agent asks in other words.

1

One move to approve

When the agent asks permission, ka allows and jaw denies. You don't need to look at the screen.

0

Screens to look at

When the agent finishes, the earbud reads its first sentence: “Fixed. 24 tests pass.”

Under the hood

It listens to muscles, not to sound.

Each move uses a different part of the tongue and jaw, so each leaves a different pattern across the sensors. A small model, trained on your own signals in a few minutes, tells them apart, and keeps adapting to you from day to day as you type.

  • Channels16 skin sensors
  • Sampling2,000 times a second
  • Signal size5 to 50 millionths of a volt
  • Where it runsOn your own computer
Muscle signal, 8 of 16 channels–
The hardware

A 16-channel biosignal lab,
the size of a matchbox.

Utter v1 circuit boardUtter v1 prototype board · 68 × 46 mm · machine-assembled
Sensing
2 × TI ADS1298 · 24-bit · 16 channels
Noise
Below 1 µV, 20–500 Hz
Link
ESP32-S3 over Wi-Fi to your computer
Safety
Battery only: recording is blocked while charging
Privacy
Recognition and the language model run locally
Cost
≈ €100 per prototype board
16channels
24-bitresolution
<1 µVnoise, 20–500 Hz
2 kHzsampling per channel
What's in the box

Every part, exactly
as it ships.

The v1 prototype kit to scale, built from the same files that make the board and print the case. Turn it, take it apart, and point at any part.

Utter v1 circuit board
Drag to turn · point attap a part
  1. 3D-printedLid

    PETG. It has openings for the electrode leads and the motion sensor, windows for the two lights, and pinholes for reset and boot. Five pegs hold the board in place.

  2. Factory-assembledUtter board

    68 × 46 mm, 4 layers, 107 parts. Two TI ADS1298 chips read 16 channels at 24-bit. An ESP32-S3 sends the signals to your computer over Wi-Fi. Charging is over USB-C, and recording stops while the cable is plugged in.

  3. IncludedBattery

    LiPo 503040: 3.7 V, 600 mAh, 40 × 30 × 5 mm, with a protection circuit and a JST-PH plug. It sits under the board.

  4. 3D-printedCase

    PETG, 79 × 62 × 17 mm closed, 91 mm across the strap lugs. Inside are a shelf for the board, rails for the battery and a pocket for its plug, plus cut-outs for USB-C and the power switch.

  5. Included4 screws

    M2 × 10 mm, self-tapping. There's no glue, so the battery can be replaced.

  6. IncludedNeckband

    Soft, adjustable fabric band with a velcro closure, 65 mm tall at the front. Its 25 mm tabs go through the case lugs, and a mark under the Adam's apple puts it in the same place every day. Washable.

  7. Included18 electrode leads

    Short snap leads from the band to the 20-pin connector through the lid: 16 sensing channels, a reference (REF) and a driven ground (RLD).

  8. IncludedDry electrodes

    Datwyler SoftPulse, reusable, on snaps sewn into the band: a 3 × 4 grid on the front of the neck, 2 on each side, and REF and RLD beside the case. Wipe them daily. A pack of gel electrodes is included for a one-time signal test.

  9. OptionalMotion sensor

    An LSM6DS3 board on 4 wires. It helps tell jaw movement apart from speech.

This is the prototype kit for early testers: the case is 3D-printed, and the band follows SilentWear, the only dry neckband with multi-day silent-speech results. The Concept view at the top of the page is the same band with the leads sewn inside.

Where we are

Built in the open,
one tested step at a time.

Done
The software

Typing, predictions and the earbud voice, tested end to end on simulated signals.

Now · Oct 2026
The hardware

Boards designed and reviewed, being ordered. Target: noise below 1 µV.

Next
Lips closed

Above 80% accuracy on the 8 moves, with lips closed.

Then
Daily use

Live typing into chats and agents, in any app.

Later
All day

The dry band worn all day. It re-adapts each morning while you type, with no calibration session.

Questions.

Does it read my thoughts?

No. It only picks up deliberate muscle activity when you make one of the eight moves. Thinking a word without moving does nothing.

Can people tell I'm using it?

Your lips stay closed and there is no sound; your tongue and jaw move slightly. Proving that this works with lips closed is exactly the next milestone.

How fast is it?

In simulation with real recognition errors from a public silent-speech dataset: about 4 words a minute on a fresh morning and 7.5 once it has adapted to you, replying to a coding agent with a local language model. A reply you sent before comes back almost free. That suits replies and agent prompts more than essays.

Which apps does it work with?

Anything you can type in: it types like a keyboard. With Claude Code it goes further, using the agent's messages as context and answering permission prompts.

Where does my data go?

From the band to your own computer over your Wi-Fi. Recognition and the language model run there, not in the cloud.

Is it a medical device?

No. It's a prototype input device, and it isn't designed or tested for medical use.

Be the first to type
without a sound.

We're building the first bands now. Join the waitlist to follow the tests and get early access.

Join the waitlist