Would be interesting to see how well their approach coped with the trivial countermeasure of playing competing recordings of (different) key sounds at the same time.

Admittedly computers can now easily spilt parallel conversations apart fairly effectively (eg cocktail party problem) but in this case with the individual noises being very similar in character it would make it harder to get the timing between genuine signals, which seems like it would make the attack much less likely to recreate an accurate key replica.