That attack is more about timing than about sound. There is some information available from the sound of a keystroke, with a very good microphone under ideal conditions each key could be identified uniquely, but that's not the main vector. Some sequences are easier to type than others by their very nature, and individual typists have idiosyncratic variance in how easily/fast/accurate they are at a given sequence. That's the main trick used to derive what's typed from a recording of it happening, and compression doesn't reduce that information, so long as the keypresses are audible.