Testing a computational model for aural detection of aircraft in ambient noise

Abstract

Computational models are used to predict the performance of human listeners for carefully specified signal and noise conditions. However, there may be substantial discrepancies between the conditions under which listeners are tested and those used for model predictions. Thus, models may predict better performance than exhibited by the listeners, or they may “fail” to capture the ability of the listener to respond to subtle stimulus conditions. This study tested a computational model devised to predict a listener's ability to detect an aircraft in various soundscapes. The model and listeners processed the same sound recordings under carefully specified testing conditions. Details of signal and masker calibration were carefully matched, and the model was tested using the same adaptive tracking paradigm. Perhaps most importantly, the behavioral results were not available to the modeler before the model predictions were presented. Recordings from three different aircraft were used as the target signals. Maskers were derived from recordings obtained at nine locations ranging from very quiet rural environments to suburban and urban settings. Overall, with a few exceptions, model predictions matched the performance of the listeners very well. Discussion focuses on those differences and possible reasons for their occurrence.