Improving PLCA-based score-informed source separation with invertible Constant-Q Transforms

Probabilistic Latent Component Analysis is a widely adopted variant of Nonnegative Matrix Factorization for the purpose of single channel audio source separation. It has seen many extensions, including incorporation of prior information derived from music scores. Recent work on the invertibility of the Constant-Q Tranform make that a viable alternative to the Short-time Fourier Transform as underlying data representation. In this paper we assess several implementations for their usability in score-informed source separation. We show that results are comparable to, and in some cases better than, use of the STFT, and that exact transform invertibility is not a significant factor in this application.