We propose a blind single-channel musical source separation method that improves perceptual quality of the separated sources. It uses the advantages of subspace learning based on Non-negative Matrix Factor 2-D Deconvolution (NMF2D). To improve the perceptual quality of separation, we propose a weighted divergence type cost function for the optimization that adopts the auditory model defined in ITU-R BS. 1387 into the source separation. It is shown that the proposed perceptually weighted NMF2D scheme efficiently clusters the bases of subspace representation corresponding to notes generated by single instruments. Source separation performance has been reported on musical mixtures resulting an improvement in perceptual quality measures.