Abstract:

We propose the method of independent subspace analysis (ISA) for separating individual audio sources from a single-channel mixture. ISA is based on independent component analysis (ICA) but relaxes the constraint that requires at least as many mixture observation signals as sources. A second extension to ICA is the use of dynamic components to represent non-stationary signals. Sources are tracked by similarity of dynamic components over small time steps. We propose a method for grouping components by partitioning a matrix of independent component cross-entropies that we call an ixegram. The ixegram measures the mutual similarities of components in an audio segment and clustering the ixegram yields the source subspaces and time trajectories. To demonstrate the techniques we give examples of ISA applied to separation of musical and speech sources from single-channel mixtures.