Frequency-Domain Blind Source Separation with Permutation Control

Peng Xie
Steven L. Grant, Missouri University of Science and Technology

This document has been relocated to

There were 17 downloads as of 27 Jun 2016.


This paper explores the problem of frequency-domain Blind Source Separation (BSS) of convolutive mixtures. The main difficulties of this approach lie in the so called permutation and amplitude problems. In order to solve the permutation ambiguity, a new hybrid approach is proposed, in which the Independent Component Analysis (ICA) processes across all frequency bins are concatenated and each of them is embedded with a permutation control unit. In each frequency bin, when the separation matrix is obtained by the ICA process, the control unit detects the possible permutation and aligns the matrix only if the permutation is confirmed. Then the final value of separation matrix is used to initialize the ICA iterations in the next frequency bin. The amplitude problem is addressed by utilizing the elements in estimated mixing matrix. The method is compared with conventional frequency-domain BSS approaches and the experimental results demonstrate superior performances of the proposed method.