Across-frequency processing in convolutive blind source separation
Jörn Anemüller · Carl von Ossiezky University of Oldenburg · 2001
Three different algorithms for the problem of separating convolutively mixed acoustic signals in the frequency domain are proposed. The first approach (chapter 2) yields an adaptive algorithm for the separation of free-field superpositions by imposing constraints on the separating filter. Separation is attained within approx. 0.2s signal time, and moving speakers are separated. The AMDecor algorithm (chapter 3) separates signals in reverberant rooms using the criterion of Amplitude Modulation Decorrelation across different frequency channels. This enables separation of the sources' spectral components and their consistent order in all frequencies in a single processing step. It is shown that separation is close to the theoretical optimum also for highly reverberant recordings. The algorithm proposed in chapter 4 is similar to the AMDecor algorithm, but uses methods from second-order statistics, making an algebraic solution possible which enables very fast separation. This algorithm has also been evaluated with other multidimensional signals (spectral image data) and is applicable to the problem of source separation with a time-varying mixing system.