Skip to content

Hands-free computing with a muscle, blink or jaw-clench switch

Short answer: if holding a key is the problem, a small, deliberate muscle contraction can stand in for it. A muscle squeeze read by an electrode can hold the microphone open like the hold-to-talk key, and you speak the words. It is experimental, it is off by default, and it has not yet been measured over real workdays, so this page is as much about its limits as its use.

Everything runs on your machine. The electrodes are read locally, no network connection is made, and the signal is not stored unless you ask for a recording. That matters for this audience: assistive tooling that depends on a service can be taken away.

Who it is for

  • People who can still contract one muscle on purpose (a forearm, a finger, the jaw or face) but cannot press or hold a key reliably, because of repetitive strain injury, a motor impairment, fatigue or pain.
  • People whose hands are busy: at a workbench, at an instrument, with a sleeping child.
  • Anyone who wants a trigger that needs no keyboard and nothing on screen.

It is not a way to type with your muscles. A wristband that decodes handwriting reaches roughly 21 words per minute in the best published result, against about 150 for speech (the research page has the numbers). So the muscle carries the intent to speak and your voice carries the words: the same split YazSes uses for every non-keyboard trigger.

What it can do

Gesture What it does by default Notes
Squeeze (a muscle electrode) A light squeeze holds the microphone open and types what you say; a hard squeeze speaks a command. The recommended trigger. Strength comes from your own calibration.
Double blink Runs a single command phrase, undo by default, through the confirmation gate. A blink is an eye-movement artifact on a frontal channel. Two quick blinks is the gesture because a single blink is involuntary.
Jaw clench Does nothing until you map it to a command. Off by default. A clench is a muscle artifact, and chewing, yawning and talking all look like one. A speech guard suppresses it while you dictate, so a clench made mid-sentence may be missed.

Blink and jaw clench are artifacts, not brain signals. On an EEG headband you are reading eye and muscle activity that leaks into the headband. The research page records the best measured blink switch at 99.5% accuracy but 0.10 false positives per minute, which is about 50 stray activations in a workday without extra safeguards. That is why the squeeze is the recommended trigger, and why blink and jaw events run only a short, reversible command rather than dictation. YazSes does not read thoughts, imagined movement or attention, and this feature does not try to.

How it works

The switch is experimental and off by default. Its steps run in this order, and each one can stop the next:

flowchart LR
    A["Electrode, headband<br/>or saved recording"] --> B["Stream check<br/>are samples arriving steadily?"]
    B --> C["Signal-quality check<br/>flat, clipped or noisy? ignored"]
    C --> D["Squeeze"]
    C --> E["Double blink"]
    C --> F["Jaw clench"]
    D --> G["False-activation budget<br/>too many cancelled? weaken or pause"]
    E --> G
    F --> G
    G --> H["Switch: start dictation<br/>or run a command"]

A stream or electrode the checks recognise as bad produces silence, not a press, and a held squeeze is released when samples stop. The checks are not perfect: in synthetic tests a large, sudden jolt can still look like a squeeze, which is why you measure your own false-activation rate before relying on it. The switch never reads thoughts: it sees only the muscle and eye activity described above.

How to start

  1. Pick the hardware. A DIY muscle sensor (MyoWare 2.0 or a BioAmp EXG Pill) on a microcontroller is the cheapest and most dependable squeeze trigger. A board BrainFlow supports (an OpenBCI board, a Muse headband) gives raw-signal control and lets YazSes measure its own mistakes; so does a device that publishes a Lab Streaming Layer stream, a NeuroSky headset, a BITalino kit or a Myo armband. The hardware guide compares them, and covers the mains setting and the safety rules. It is not a medical device, and electrodes must never be connected to mains-powered equipment.
  2. Try it without hardware. Set source = "synthetic" and run yazses biosignal monitor to see scripted gestures go through the real detectors.
  3. Check the real signal. yazses biosignal doctor checks the chain and ends with a Verdict: line telling you whether you are ready or what to fix first. Then yazses biosignal calibrate learns your relaxed level, a light and a hard squeeze, a blink and a clench. yazses biosignal evaluate scores the detectors on a scripted session or on a recording you labelled.
  4. Measure how often it fires by mistake. yazses biosignal field-test takes about ten minutes: you sit, talk, chew, type and look around with the switch live (nothing is typed), then make each gesture on cue. You get the phantom activations per hour for each activity, with an upper bound. Repeat it on another day.
  5. Use it, and read your own numbers. yazses biosignal report shows how often each gesture fired and how often you cancelled it. That is the evidence that decides whether you keep it on.

The step-by-step configuration is in Use a muscle, blink or jaw-clench switch.

A switch that watches itself

A trigger that fires when you did not mean it is worse than none. Three guards keep that in check:

  • The false-positive budget. YazSes counts the activations you cancel. If the rate goes over the budget you set, it first desensitizes the detectors, and if the cancelled rate stays over the budget it then suspends them, and you release it with one action (yazses handsfree resume). A suspension also lifts by itself after a timer, so it is never permanent.
  • A calibration that has to fit. A saved calibration made for a different board, different channels or a different sampling rate, or one older than a month by default, is not used. The muscle-squeeze switch stays off until you recalibrate, rather than guessing.
  • A global pause. With the hands-free global stop on, yazses handsfree pause stops the switch with the other hands-free inputs, so you can stop it from a terminal or the settings window.

Honest limits

  • Unmeasured. The detectors have been tested on synthetic and replayed signals. They have not been shown to work for any particular person, electrode or board, and every default is a provisional starting point. No false-activation rate has been measured on a real person yet, by this project, and the research page records that the best-known wristband paper reports none either: the field test above is how you get the number for yourself, and it is only a number about you.
  • It is personal. Skin, electrode placement and muscle differ between people. Calibrate to yourself, and again whenever an electrode moves.
  • A bad signal means no events. A flat, clipped or hum-filled channel is ignored on purpose, and the tools say why.
  • Fatigue and spasm are yours to judge. A trigger that needs a repeated squeeze can itself tire you. If it does, lower the demand (a lighter squeeze, a different muscle) or use another trigger.
  • Not a medical device. It is not for diagnosis, monitoring or safety-critical control.

If you try it, please report what worked and what did not: accessibility feedback is the most useful feedback this project can receive, and an issue is the place for it. Do not attach a recording of your signal to a public issue; share the output of yazses biosignal doctor instead.