Performance of independent component analysis when used to separate competing acoustic sources in anechoic and reverberant conditions

A review of existing methods for independent component analysis was carried out and a series of experiments conducted assessing the use of existing independent component analysis (ICA) methods to separate microphone sources in varied acoustic environments. Specifically the research looked at how effectively ICA could perform in a broadcast context using standard microphone techniques such as spaced omni and coincident crossed cardioid pairs. Experiments were carried out in an anechoic chamber and also in a listening room conforming to the ITU-R BS.1116-2 standard. Comparisons showed a large variance in the performance of different ICA algorithms and results clearly indicate the limitations of ICA when performed on audio material recorded in a reverberant environment however it was still shown possible to achieve separation of signals of up to 12dB even in these conditions