Combining Second-Order Statistics BSS and LCMV Beamforming
Adaptive LCMV beamforming and convolutive blind source separation (BSS) have a common goal, namely to reduce interferences. On the one hand, BSS algorithms are able to adapt continuously, while LCMV beamforming algorithms adapt only when the interferer signal is dominant. On the other hand, adaptive LCMV beamforming algorithms may converge faster than NG-SOS-BSS algorithms if there is no double-talk, as we have observed in Chap. 8. Moreover, beamformers may exploit geometric prior information about the position of the target source. In this chapter our objective is to combine the advantages of both approaches.
A combination of a BSS algorithm and a beamforming algorithm depends on several design options. At first glance, the beamformer could be either data-independent or adaptive, or it could be placed either in front of or after the BSS block, leading to four possible combinations. However, two combinations may be removed a priori. Firstly, placing a data-independent beamformer after the BSS block is not appropriate. This is because the BSS separation system distorts the spatial properties of the signals on which data-independent beamformers are based. Secondly, the combination of an adaptive beamformer in front of the BSS block can also be removed a priori. The adaptive LCMV beamformer is generally highly time varying. The BSS separation system could bring an improvement if it was able to track the time variance of the beam-former, that is, if it converged faster than the beamformer. This is generally not the case, as shown in Chap. 8.
KeywordsAutomatic Speech Recognition Separation System Blind Source Separation Soft Constraint Interference Canceler
Unable to display preview. Download preview PDF.