Blind Source Separation with Non-stationary Mixing Using Wavelets

This paper addresses the problem of blind source separation in the situation where the mixing process is dynamic. We first present a new ICA algorithm for the static mixing problem that exploits a wavelet representation of the signals. This outperforms standard ICA in our experiments thus allowing the unmixing to be estimated from a smaller number of samples. We use this to create a sliding window based algorithm that is capable of tracking the dynamics a non-stationary mixing process in the blind source separation problem. An effective initialization for each new window is calculated based on a smoothed estimate of the unmixing process learnt in previous windows. This reduces the computation required for updating each window and reduces the chance of the algorithm falling into undesirable local minima of the cost function. The efficacy of the algorithm is demonstrated on some simulated data using artificially mixed audio sources.