Testing a computational model for detection of “real-world” sounds

Abstract

Hoglund (et al., 2010) reported the ability of listeners to detect a recorded signal masked by nine different ambient sounds. Signals were selected from 410 recordings of two different helicopters; maskers were selected from nine outdoor field recordings from three different locations. Signal and masker levels were randomized from trial to trial in a 2I,2AFC adaptive procedure to produce psychometric functions. Signal-to-masker ratios (SNR) ranged from -22 to -4 dBA for 70% correct detections. The Hoglund et al. results were used to tune a network of non-linear oscillators to optimize its performance. To evaluate the model, recordings of three fixed-wing aircraft were obtained for use as signals in the same nine environmental backgrounds. Signal and masker recordings were pre-processed to ensure that overall RMS voltage measures reflected the sound pressure level measured in a flat-plate coupler at the headphone used by the listeners. Model predictions for complete psychometric functions were obtained before experimental results were available for comparison. This poster presents details of the pre-processing of signals and maskers, and of the adaptive, single-interval procedure used to obtain the SNR for P(C) = 66.5%, 75%, and 87.5% by 16 listeners. Good agreement was found between model predictions and listener performance. [Work supported by a SBIR grant from AFOSR.]