Salience: Dynamics Processing in the Information Domain

SALIENCE is a real-time spectral audio effect whose controlled quantity is not a property of the signal but a property of a model listener's beliefs: how surprised that listener is. The listener hears the effect's output and learns from it, so the effect closes a loop around a learning system. For any predictor that learns at a constant rate, the total signed prediction error caused by a change of size D is exactly D divided by the learning rate, whatever way the change is revealed. Processing can therefore move that error in time and reshape it, but cannot remove it without changing what is finally learned. This is a property of constant-gain predictors, not of surprise in general: information-theoretic surprise obeys no such law, and the spread-normalised error that SALIENCE regulates obeys it only approximately. SALIENCE builds five controls on this: a target surprise (FOCUS), a surprise quantizer that holds a change starting off the beat grid and releases it on the next line (GRID), and a 0.25 s lookahead that acts on the listener's expectation before a change arrives (PRIME), a separation of lasting changes from one-off events by whether their signed error totals to zero (MEANING), and a ceiling on how much new the listener receives at once (RATE). Only what is new against the music's structure is touched: in steady state, after a two-bar warm-up, a loop that repeats sample-exactly leaves the output equal to the delayed input to the floating-point floor at every setting tested, and kick and bass are protected. The exception is a transient: at calm FOCUS, an afterimage formed during the warm-up fades over about 20 s (below −60 dB after 4 bars) before the floor is reached. Humanised timing is not passed exactly: with 5 ms of random timing jitter per hit, 2 % of band-frames are touched at the defaults, by up to 15.5 dB. Synthetic benches show the quantizer holding a new lead at −13.6 dB until the bar line and releasing it at full level, and priming lowering the model listener's surprise at a lead's arrival from 10.9 to 6.9 standard deviations while its first-beat level is turned down less (−3.8 dB instead of −5.2 dB). An audit of all eleven controls through the plugin itself passes 15 of 15 checks; the conclusions hold when seven of the listener's nine constants are moved by half either way; and an independent surprisal model agrees on the direction of priming and quantization. All claims about surprise concern the model listener; the note includes a registered listening-test protocol (Section 9) that has not yet been run.