Working on a real-time audio classifier recently, I keep noticing how the gate-before-processing pattern from this era still rules modern audio ML. Feed silence into a sound classifier and it'll happily hallucinate something — so you put a noise gate on the input, exactly like the trick described here, just used for a different purpose.

Counterintuitive thing I learned: when I tried to skip the explicit gate and 'let the model learn it', accuracy dropped meaningfully. The deterministic preprocessing wins over end-to-end here. Kind of an inverse of Padgham's 'accident becomes intent' — the technique survives, just on the analysis side now instead of the production side.