Goal-Oriented Auditory Scene Recognition

摘要

How do we understand and interpret complex auditory environments in a way that may depend on some stated goals or intentions? Here, we propose a framework that provides a detailed analysis of the spectrotemporal modulations in the acoustic signal, augmented with a discriminative classifier using multilayer perceptrons. We show that such representation is successful at capturing the non-trivial commonalties within a sound class and differences between different classes. It not only surpasses performance of current systems in the literature by about 21%, but proves quite robust for processing multi-source cases. In addition, we test the role of feature re-weighting in improving feature selectivity and signal-to-noise ratio in the direction of a sound class of interest.