A novel multimodal approach for independent component analy-rnsis (ICA) of complex valued frequency domain signals is presentedrnwhich utilizes video information to provide geometrical descriptionrnof both the speakers and the microphones. This geometric informa-rntion, the visual aspect, is incorporated into the initialization of therncomplex ICA algorithm for each frequency bin, as such, the methodrnis multimodal since two signal modalities, speech and video, arernexploited. The separation results show a significant improvementrnover traditional frequency domain convolutive blind source separa-rntion (BSS) systems. Importantly, the inherent permutation problemrnin the frequency domain BSS (complex valued signals) with the im-rnprovement in the rate of convergence, for static sources, is shownrnto be solved by simulation results at the level of each frequency bin.rnWe also highlight that certain fixed point algorithms proposed byrnHyv ¨ arinen et. Al., or their constrained versions, are not valid forrncomplex valued signals.
展开▼