Multichannel Nonnegative Matrix Factorization With Motor Data-Regularized Activations For Robust Ego-Noise Suppression

The suppression of ego-noise is often addressed using dictionary-based methods where the characteristic spectral structure of ego-noise is approximated by a linear combination of dictionary entries. A blind, entirely audio data-based selection of the dictionary entries is, however, challenging and reacts sensitive against other signals besides ego-noise in a mixture. For a more robust behavior, we propose a motor data-dependent regularization term which promotes similar activations for similar physical states of the robot. The proposed regularization term is added to a multichannel nonnegative matrix factorization (MNMF)-based signal model and according update rules are derived. We analyze the proposed method for a challenging ego-noise scenario and demonstrate the efficacy of the method compared to an approach for which no motor data is used.