Adaptive unsupervised separation of discrete sources

This paper treats source separation with the help of contrast functions and proposes corresponding adaptive implementations. Its major contributions are two-fold: (i) it proposes a new contrast which can be evaluated without pre-whitening the signals, provided all have unitary power. Its adaptive maximization involves an output AGC for each source recovery and performs as well as separation with pre-whitened signals; (ii) in case of sources with discrete alphabet, an intermediate contrast is proposed which takes additional advantage of the alphabet knowledge. The improvement to source separation is significant for correlated signals, but for adaptively pre-whitened separation, the quality of whitening conditions the improvement. ( 1999 Elsevier Science B.V. All rights reserved.