A GPU-accelerated real-time implementation of TRINICON-BSS for multiple separation units

In this demonstration, a GPU-accelerated implementation of the second-order statistics version of the TRINICON (Triple-N independent component analysis for convolutive mixtures) blind source separation (BSS) algorithm is presented. The single-precision arithmetic separation performance is compared with a double-precision CPU (Central Processing Unit) reference implementation to establish the effect of parallel algorithms and the reduction of floating-point precision. Computational performance results are presented for a single 2-channel separation unit and multiple simultaneous 2-channel separation units to study the potential future use in a sensor array network with multiple 2-channel separation units.