Separating Short Signals in Highly Reverberant Environment by a Recursive Frequency-Domain BSS

A new approach to the permutation problem for Blind Source Separation (BSS) in the frequency domain is presented. The independence of the separation across the frequencies, and thus the probability that a permutation may occur, is minimized by a recursive linking of the ICA stage. A recursive adaptive estimation of smooth demixing matrices is used to initialize the Independent Component Analysis (ICA) in order to force it to converge with a coherent permutation across the whole spectrum. Since no information about non stationarity of the signals is exploited, the proposed method works also for short utterances (0.5-1 s) and in highly reverberant environments (T₆₀sime 700 ms). Furthermore it is shown that the recursive initialization increases the accuracy of the ICA when a small amount of data observations is available.