Markov Keyboard: keyboard layout that changes by Markov frequency (2019)

Nice. Some thoughts.

Look at PPM. Your prediction model would work better with personalised data. PPM is efficient (nb. I see you are using python - look at this https://github.com/willwade/pylm - although be warned - i think my code is not quite right..)

Layout shifting for finger movement - well its great if you didnt have to look. The time for visual processing the letters adds a significant lag (its why typical word prediction isnt used that much and when it is - not over 3 predictions (I have papers on this if you are interested). But its not all bad..

Switch users who need next letter prediction this could dramatically support their rate of input. (view https://youtu.be/Bhj5vs9P5cw?si=VnytfH_vdEUWuLok&t=73 - now note how the keyboard blocks the scan up. But imagine if it just scanned each letter first by next most likely - or heck - like this repo - actually changes button position and kept the scan pattern the same. It would be a ton more efficient)

(and a bit of a rabbit hole.. What if keys had word predictions on them? This is basically the end result of ACE-LP: https://discovery.dundee.ac.uk/en/publications/ace-lp-augmen...)

This is not a bad idea. It's not dissimilar from an autocomplete engine. If reset per word and using a static probability table it will work its way into muscle memory. In practice fingers will only need to leave the home row for some initial letters.

Fun concept. I had a thought about having it optimize in a way where left and right hands alternate for increased speed

I think that this is an attempt to stop the “clean key” problem. Security system keypads (especially outdoor ones) tend not to be cleaned, so as time passes, it is easy to spot the dirty keys. Dirty keys are not being pressed and are therefore not in the passcode.

So look for the clean keys and try combinations from there. In a 4 digit (0-9) keypad, knowing the clean numbers drops the possible codes from 9999 to 24 (if my early morning math holds up).

Also helps the issue of someone looking over the shoulder of a valid person. Chances are they are just seeing the position and not the character pressed. So the keyboard changes and you actually need to know the character not just the old position.

Too bad this github has latest changes 5 years ago, I wonder what they could update with LLMs, etc.