Detection of early reflections from a binaural activity map using neural networks

Human listeners localize sounds to their sources despite competing directional cues from early room reflections. Binaural activity maps computed from a running signal can provide useful information about the presence of room reflections, but must be inspected visually to estimate auditory cues. A model was constructed using machine learning to validate the presence of and perform the extraction of these cues. The model uses the running signal output of a binaurally integrated cross-correlation/autocorrelation mechanism (BICAM) to analyze a lead/lag stimulus and generate a binaural activity map. System reflections are visually presented on the binaural display as correlation peaks with increased amplitude. Three independent neural networks estimate the location of the direct sound, the time delay of the reflection, and the location of the reflection from binaural activity maps displayed by BICAM. Depending on the task, neural network accuracies on test data sets vary from 84.1% to 98.5%.