Single Channel Blind Source Separation Using Independent Subspace Analysis

Single Channel Blind Source Separation Using Independent Subspace Analysis by Jason Heeris Supervisor: Dr. Roberto Togneri The problem of separating conceptually distinct sources of information in a single channel mixture signal, known as single channel blind source separation, was approached using the technique of independent subspace analysis, an extension of independent component analysis. A prototype system was implemented and tested in the numerical processing language Octave and showed reasonable success at separating simple test signals. The prototype failed to adequately separate mixtures of speech and noise, however, and its performance was severely degraded when adapted to operate on non-stationary signals. The inability to select an optimal level of detail to retain during processing coupled with the unsatisfactory non-stationary operation appear to be the main weaknesses of this technique, and further development should focus on improving these points.