Investigating computational models of perceptual attack time

The perceptual attack time (PAT) is the compensation for differing attack components of sounds, in the case of seeking a perceptually isochronous presentation of sounds. It has applications in scheduling and is related to, but not necessarily the same as, the moment of perceptual onset. This paper describes a computational investigation of PAT over a set of 25 synthesised stimuli, and a larger database of 100 sounds equally divided into synthesised and ecological. Ground truth PATs for modeling were obtained by the alternating presentation paradigm, where subjects adjusted the relative start time of a reference click and the sound to be judged. Whilst fitting experimental data from the 25 sound set was plausible, difficulties with existing models were found in the case of the larger test set. A pragmatic solution was obtained using a neural net architecture. In general, learnt schema of sound classification may be implicated in resolving the multiple detection cues evoked by complex sounds.