I think this linited attack surface can work without having to generalize one model to multiple people or keyboards. One advantage of a Zoom attack is that you get “plaintext” shortly after hearing the “ciphertext” if you can get the target to type into the chat window. And when you hear typing in other contexts it’s likely to be something that matches a handful of grammars that an LLM can recognize already (written languages, programming languages, commands, calculation inputs) - and when it doesn’t, that’s probably a password.