Robuston methods for stable statistical signal processing: principles and application to nonstationary signal estimation

We introduce a reduced-detail paradigm for nonstationary statistical signal processing with enhanced performance. Time-frequency localized subspace signal components (called robustons) are used as atomic entities for statistical signal modeling and processing. Robuston signal processing employs special time-varying filters that allow an efficient on-line implementation, and statistical signal descriptors that can be estimated in a stable manner by means of intra-subspace averaging. We develop the principles of robuston signal processing and consider optimal nonstationary signal estimation as a specific application. The performance advantages of the resulting "robuston Wiener filters" are assessed by means of simulations.