Insights Into Deep Non-Linear Filters for Improved Multi-Channel Speech Enhancement

The key advantage of using multiple microphones for speech enhancement is that spatial filtering can be used to complement the tempo-spectral processing. In a traditional setting, linear (beamforming) and single-channel post-filtering are commonly performed separately. contrast, there trend towards employing (DNNs) learn joint non-linear filter, which means restriction processing model separate information potentially overcome. However, internal mechanisms lead good performance such data-driven filters multi-channel not well understood. Therefore, in this work, we analyse properties filter realized by DNN as its interdependency with temporal spectral carefully controlling sources (spatial, spectral, temporal) available network. We confirm superiority model, outperforms an oracle challenging speaker extraction scenario low number 0.24 POLQA score. Our analyses reveal particular should processed jointly increases selectivity filter. systematic evaluation then leads simple network architecture, state-of-the-art architectures on task 0.22 score 0.32 CHiME3 data.