Multimodal Dictionary-based Ego-Noise Suppression for Acoustic Self-Awareness of Autonomous Systems

In this thesis, ego-noise suppression for autonomous systems is considered. It is proposed to model ego-noise using so-called dictionaries which will turn out to be especially suited to represent the spatial and spectral characteristics of ego-noise. Specifically, semi-supervised single- and multichannel nonnegative matrix factorization (NMF) are introduced for ego-noise suppression. Furthermore, it will be shown that ego-noise suppression can benefit significantly if motor data, i. e., angle information collected by proprioceptors of joints and motors, is included in the suppression algorithms.