Impact of Coding Noise on the Convergence of Blind Source Separation

In networks consisting of distributed microphones, signals from several spatially distant microphones are exploited for multichannel signal processing. This requires transmission of at least one microphone signal to another network node. In this paper, the effect of a predictive source coding scheme on the TRINICON-based Blind Source Separation (BSS) algorithm is evaluated. First, the propagation of the quantization noise into the separated output signals is analyzed and verified experimentally. Moreover, the effect of coding on the convergence of the algorithm itself is evaluated. For the experimental part of the analysis, the application of BSS in a binaural hearing aid scenario is considered. BSS proves to be very robust against small coding word lengths and exhibits a minor sensitivity on noise shaping.