Effects of Audio Coding on ICA Performance: an Experimental Study

In this paper, we study the influence of lossy audio coding on the performance of Independent Component Analysis (ICA). In particular, we derive two compression scenarios from practical implementations. In the first case, we consider the situation when the sources are independently compressed, decompressed and then mixed. In the second, we consider the situation when mixtures of sources are jointly compressed. We experimentally show that the modification of the spectrotemporal diversity due to compression has almost no effect on the performance of ICA methods. We also show that the two tested stereo encoding strategies have a major effect on the performance of ICA, especially when the mixed signals are compressed at low bit rates. As these strategies have been extended to audio systems involving much more than two channels, our work suggests that ICA will not be able to successfully process the high number of observations provided by modern immersive audio systems.