All purely conjecture, but I'm fairly convinced this method of audio analysis fingerprinting does not work in real life. In the ocean for subs yes but keys probably not.
The initial attack of the audio signal from a mechanical keyboard is pretty loud, so it would be fairly difficult to mask.Systems that mask 40db SPL to 80db SPL for ambient sounds would not do a good enough job of blocking that sudden spike. But rubber knob keybeds might offer better mitigation.
If the technique of measurement is through drops in volume across distance, simply shifting the distance or rather the position of your keys so as not to get a clusterable audio of your keyboard would spoil the sampling stage for training data.
The idea here is to flummox the initial data gathering stage so there isn't a ground truth to base any further analysis on. Obviously the more tinfoil hatters have multiple keyboards lying around with which they randomly type strings for precisely this purpose.
Also if they are using particular characteristics of the keys themselves, then what I previously pointed out with switching keymaps don't work. Then pre sampling the keys would work.
The other way, and this is probably feasible, would be to use magnetic field detection, maybe the phone compass works this way, to detect the key code signals as it goes down the keyboard wire.
In any case it is far easier to install a hidden camera to record keystrokes than it is to do this audio analysis.