Multichannel Nonnegative Matrix Factorization for Ego-Noise Suppression

A mobile robot generates significant self-induced noise, e.g., by its joints and moving parts of its body, which is referred to as ego-noise. Ego-noise is of crucial importance in robot audition since it massively corrupts the microphone signals and thus degrades the robot's capability to interact intuitively with its acoustic environment. Therefore, appropriate mechanisms for ego-noise suppression are required. In this paper, we investigate different Multichannel Nonnegative Matrix Factorization (MNMF) algorithms with respect to their suitability for ego-noise suppression, propose a modified semi-supervised MNMF approach and evaluate its performance.