A multimodal approach for frequency domain independent component analysis with geometrically-based initialization

Abstract

A novel multimodal approach for independent component analysis (ICA) of complex valued frequency domain signals is presented which utilizes video information to provide geometrical description of both the speakers and the microphones. This geometric information, the visual aspect, is incorporated into the initialization of the complex ICA algorithm for each frequency bin, as such, the method is multimodal since two signal modalities, speech and video, are exploited. The separation results show a significant improvement over traditional frequency domain convolutive blind source separation (BSS) systems. Importantly, the inherent permutation problem in the frequency domain BSS (complex valued signals) with the improvement in the rate of convergence, for static sources, is shown to be solved by simulation results at the level of each frequency bin. We also highlight that certain fixed point algorithms proposed by Hyvärinen et. al., or their constrained versions, are not valid for complex valued signals.