From Time-domain Separation of Stationary Temporally Correlated Sources to Frequency-domain Separation of Nonstationary Sources

This paper demonstrates and exploits some interesting frequency-domain properties of nonstationary signals. Considering these properties, a new method for blind separation of linear instantaneous mixtures of mutually uncorrelated, nonstationary, real sources is proposed which is based on spectral decorrelation of the sources. It allows the existing timedomain algorithms developed for stationary, temporally correlated sources to be applied to nonstationary, temporally uncorrelated sources just by mapping the mixtures into the frequency domain. The method sets no constraint on the piecewise stationarity of the sources, unlike most of previously reported methods.