A Sparse Non-Parametric Approach for Single Channel Separation of Known Sounds