Sound Source Separation of N Sources from Stereo Signals via Fitting to N Models Each Lacking One Source

We present a system to perform sound source separation of an arbitrary number of speech or music sources from a stereo signal. We build on the work of other authors’ DUET system, which uses a histogram technique to estimate the mixing parameters of the time-frequency sparse sources, before using a nearest-neighbor approach to demix the sources. Herein, we describe a new demixing method called Delay and Scale Subtraction Scoring (DASSS) that is less erratic than the nearest-neighbor method, and highlights when the sparsity assumptions of the DUET system are not valid. We also utilize a demixing technique to be used in cases where multiple sources are present, and propose an additional source-aware demixing technique for such cases. We demonstrate psychoacoustically convincing results on an example signal.