Lip phrases Beta 0.1
Teach MouthRead a short list of your own phrases by silently mouthing each one 2 or 3 times. Then mouth a phrase and it shows up in big letters, with a button to say it out loud.
- Works best after training, for a small set of phrases (start with 5 to 10), with the same person, camera and lighting you trained with. It only knows the phrases you trained; it does not read any other words.
- Not a medical device. Not for emergencies or critical medical decisions. In an emergency call 911 or use the nurse call button. By using it you agree to the terms.
- Stays on this device. The camera picture is never recorded or uploaded. Only lip-shape measurements for the phrases you train are saved, in this browser. Privacy
Lip phrases needs JavaScript turned on.
Templates belong to the person who trained them. If you point the back camera at someone else, they should train their own phrases. Only point the camera at people who know and agree.
Or did you mean…
Tip: start and end with your lips closed and still for a moment.
Pick a phrase, press Record take, and silently mouth it once at your normal pace. Do 2 or 3 takes per phrase. Start with the 5 to 10 phrases you use most.
Check my training
Tests your takes against each other. Phrases that look alike on the lips (for example "I'm hot" and "I'm hungry") get mixed up; retrain or reword one of them.
Phrases
Short, distinct phrases work best. Phrases that look the same on the lips can't be told apart.
Settings
Your training data
Saved only in this browser on this device. Export it to keep a backup or move it to another device (it contains lip-shape numbers, not video).
Debug (match distances)
Lower distance = closer match. "not sure" fires above the threshold. Values are from the last reading.
No data yet.
| Phrase | Dist | Conf | Per-take distances |
|---|
How it works, honestly
When you train, MouthRead uses Google's open-source MediaPipe face tracker, running inside this browser, to measure your lips about 30 times a second: how open, how wide, the corners, the inner-lip area and how fast they change. Each take is saved as a short list of those numbers.
When you read, it waits for your lips to move and stop, then compares the movement with every take you trained (a method called dynamic time warping, which allows for speaking faster or slower) and shows the closest phrase. It is matching, not true lip reading: it can only pick from your list, and phrases that look alike on the lips (p, b and m all look the same) will get confused.
Your phrases are matched with our own code. No lip-reading AI model is downloaded or used. See licenses.