They're using a math technique similar to how audio fingerprinting (e.g. can identify a song from someone humming it) works. Also, I could be wrong but it might actually be a clever way to use the built-in gesture recognition features of some of the newer wifi 7 chipsets (yeah, they have that).

They're not using it to pick up audio (obviously), but the way the math works would be similar: When someone moves their hand this way, that's a 90% match for the letter A. That sort of thing. So when you capture a series of movements, you can associate that with the person typing a specific word.

What I'd like to know is how specific this technique is to the keyboard/person in question. As in, if it was trained on the researchers/volunteers and thus only works on the researchers/volunteers.

I suspect that it's likely got the same problem as keyboard audio fingerprinting tech: It only works if you train it ahead of time for that user on that specific keyboard.