Deriving Spectro-temporal Properties of Hearing from Speech Data

Human hearing and human speech are intrinsically tied together, asthe properties of speech almost certainly developed in order to beheard by human ears. As a result of this connection, it has beenshown that certain properties of human hearing are mimicked withindata-driven systems that are trained to understand human speech.In this paper, we further explore this phenomenon by measuring thespectro-temporal responses of data-derived filters in a front-end convolutionallayer of a deep network trained to classify the phonemesof clean speech. The analyses show that the filters do indeed exhibitspectro-temporal responses similar to those measured in mammals,and also that the filters exhibit an additional level of frequency selectivity,similar to the processing pipeline assumed within the ArticulationIndex.