Mixed-signal real-time adaptive blind source separation

A mixed-signal adaptive VLSI architecture for real-time blind separation of linear source mixtures is presented. The architecture is digitally reconfigurable and implements a general class of independent component analysis (ICA) update rules in common outer-product form. In conjunction with gradient flow, a technique for converting wave sources into equivalent linear instantaneous mixtures by observing spatial and temporal derivatives of the field over a miniature array, the ICA architecture allows to separate and localize multiple acoustic sources in the acoustic scene. Experimental results from VLSI implementation of the ICA architecture demonstrate 30 dB separation of two mixtures of two speech signals.