Interpreting End-to-End Deep Learning Models for Speech Source Localization Using Layer-wise Relevance Propagation

The networks learn to denoise and dereverberate the microphone signals to better correlate them and consequently estimate the source position.