Abstract
The blind source separation (BSS) problem begins when multiple source signals are mixed together and then observed. The fundamental goal of BSS is to de-mix and estimate those signals using only the observations, without any knowledge of how the signals were combined and without any knowledge of the underlying signals. The field of BSS is over thirty years old and, although there exist a multitude of BSS methods that aim to solve the problem, each method is limited by the assumptions made about how the signals were combined or about the signals themselves. The current work proposes to use multiple BSS algorithms running in parallel to enhance the BSS solution and make it more robust without requiring knowledge of the signals or the mixing system.In this work, the proposed multiple algorithm source separation (MASS)approach is outlined and is implemented in a software framework. Results demonstrate that MASS provides as good or better source separation than any of the single BSS methods considered under a variety of conditions. Although BSS is the goal, the more general (not just blind) source separation problem is the basis for the framework. This allows the study/use of both blind and supervised methods, as well as source extraction methods (methods that isolate a single signal) and allows them to interact.The MASS framework presented here also offers capabilities to simplify future BSS research, allowing researchers to compare, communicate and reproduce results.These features also are generalizable to encompass other, non-BSS, multiple algorithm signal processing domains.