I'm also super excited to see more on this, it's been living rent free in my head for a long time. I played around with keypress extraction from sound recordings a couple decades ago in college, and it was remarkably straightforward to extract keypresses based on sound signals, even with zero training - based on the time between keypresses and the unique signature of space/return/backspace you can build a predictor pretty quickly. We never made it to the "bounce a pair of lasers off two window panes a known distance apart and triangulate everyone's keyboard", but it was one of those things that's definitely doable with enough time and brains.
I remember making a demo in 6th or 7th grade of console login dialog and disclosing password asking friends to log in.. they couldn't, because during typing my program would analyze time between pressing different keys.. I don't remember if I hard coded it or it was trained first... But yeah, essentially it was checking not just the password, but also signature of it..
