Could you maybe do away with training beforehand on a single source and instead use multiple, very sensitive microphones and triangulate the locations of the ke

Could you maybe do away with training beforehand on a single source and instead use multiple, very sensitive microphones and triangulate the locations of the keys being pressed? The estimations might not be accurate, but you could put the results through a smartphone typo correction algorithm.