Perceptibility Ablation
In this section, we demonstrate the necessity of our psychoacoustic perceptibility constraint by comparing samples with similar attack success (all produce STOI degradation by the model of approximately -1.1; not including an offset produces higher degradation of -1.4, but is much more perceptible; see paper for details). While different attacks produce perceptually different outcomes in the model, our emphasis here is on the perceptibility of the attack: we believe that different constraints from ours produce more perceptible perturbations.
Original Audio
Initial Model Output
Attacked Audio
Attacked Model Output
Attacked Audio
Attacked Model Output
Attacked Audio
Attacked Model Output
Attacked Audio
Attacked Model Output
Attacked Audio
Attacked Model Output