Shift-invariant Sparse Coding for Single Channel Blind Source Separation

In this paper we present results on single channel blind source separation based on a shift-invariant sparse coding model [1], [2] and [3]. This model learns a set of time-domain features from a single observation of the mixed signals. The found features can often be associated with a single source and can therefore be used to reconstruct the individual source signals. This is shown in this paper on two real world examples, the separation of fetal and maternal heartbeats from a single ECG recording and the separation of singing and accompanying guitar from a musical recording. In the first problem we learn two features, one representing the fetal heartbeat and one representing the maternal heartbeat. In the second example we learn a much larger set to model the more complex source signals and therefore introduce a clustering method to associate the different features with each of the sources.