Minimal Distortion Principle versus Back Projection for Independent Vector Analysis

This short communication deals with the scaling ambiguity issue in blind convolutive source separation when performed in the frequency domain. It discusses the relationship between two major techniques, mainly the Minimal Distortion Principle and the Back Projection, that allow to overcome the aforementioned indeterminacy. The Minimal Distortion Principle minimizes the mean square difference between the separated sources and the sensor signals, while the back projection recovers the sensor-observed amplitudes of each estimated source signal. Herein, we prove that the Minimal Distortion Principle is a particular solution of the Back Projection. Another contribution of this paper consists of exploiting one of the most beneficial outcomes of the Back Projection, that is spatial diversity. Our proposed approach applies Single Input Multiple Output deconvolution to the outputs of the back projected source signals, after their estimation by the Independent Vector Analysis algorithm. This method has the advantage of improving the estimation accuracy and removing the channel effect. Experimental results show the effectiveness of our proposal with respect to both the Minimum Distortion Principle and the conventional Back Projection solution.