Deep learning of the binaural masking level difference

Abstract

The binaural masking level difference (BMLD) is the improvement in the detection of a signal in noise observed for different interaural configurations. Durlach [J. Acoust. Sci. Am. 1206–1218 (1963)] derived equations that accurately predict much of the human BMLD psychophysical data. We trained a deep neural network to predict BMLDs, as calculated with Durlach's equations, based on waveforms of a 500 Hz signal in a white noise each presented from different locations. Crucially, the network was constrained via nodes configured to embody informative representations (Iten etal., 2018; arXiv:1807.10300). The deep neural network accurately predicted BMLDs for stimuli within a simulated azimuth and for stimuli with interaural time differences (ITDs) outside the range of a human head. Further, even though the model was not designed to imitate neural biophysics, we discovered that the dynamics of latent nodes bore similarities with published data on neural ITD tuning and rate-level responses to BMLD stimuli. The work demonstrates how advances in deep learning can be used to consolidate theoretical and experimental approaches to binaural detection.