A comparison of decorrelation criteria for the blind source separation of nonstationary signals

We compare and contrast two frequency-domain decorrelation criteria for the blind source separation of convolutive mixtures of nonstationary signals. Both criteria operate on the coherence function matrix measured on the outputs obtained by filtering and combining the sensor signals. One criteria involves the Frobenius norm of the coherence function matrix, which is related to the sums of squared coherence functions between all possible pairs of outputs. The other criteria is the determinant of the coherence function matrix, which is related to the volume of state space occupied by the outputs. These criteria are identical for two outputs, but differ for three or more outputs. Full gradient descent of either criteria involves simultaneously decorrelating the outputs while maximizing the output energy. Thus, we will also compare and contrast the resulting algorithms to decorrelation alone. Quantitative comparison will be done using data recorded in a real acoustic environment.