Single-channel source separation problems occur when a number of sources emit signals that are mixed and recorded by a single sensor, and we are interested in estimating the original source signals based on the recorded mixture. This problem, which occurs in many sciences, is inherently under-determined and its solution relies on making appropriate assumptions concerning the sources. This dissertation is concerned with model-based probabilistic single-channel source separation based on non-negative matrix factorization, and consists of two parts: i) three introductory chapters and ii) five published papers. The first part introduces the single-channel source separation problem as well as non-negative matrix factorization and provides a comprehensive review of existing approaches, applications, and practical algorithms. This serves to provide context for the second part, the published papers, in which a number of methods for single-channel source separation based on non-negative matrix factorization are presented. In the papers, the methods are applied to separating audio signals such as speech and musical instruments and separating different types of tissue in chemical shift imaging.
Main Research Area:
Technical University of Denmark, DTU Informatics, Building 321, 2009