You'd be amazed what can be done with the latest machine learning algorithms. But setting those aside, you could still get 90% of the way with some relatively basic stuff.
I'd start by identifying the presence of speech—when the user isn't currently speaking, trivial muting is all that's needed. Then the only challenge would be to squelch keyboard noise that overlaps speech. If keyboard input monitoring has enough temporal precision, even the most trivial volume dip (and perhaps a band pass filter) would be a win.
Keyboard input monitoring could also be used to continuously tune the algorithm to the loudness of the keyboard noise. Any key press occurring in sonic isolation could be used. Then when there's keyboard noise which overlaps speech, the algorithm will know how much to squelch.
I'd start by identifying the presence of speech—when the user isn't currently speaking, trivial muting is all that's needed. Then the only challenge would be to squelch keyboard noise that overlaps speech. If keyboard input monitoring has enough temporal precision, even the most trivial volume dip (and perhaps a band pass filter) would be a win.
Keyboard input monitoring could also be used to continuously tune the algorithm to the loudness of the keyboard noise. Any key press occurring in sonic isolation could be used. Then when there's keyboard noise which overlaps speech, the algorithm will know how much to squelch.