A Musical Source Separation System Using a Source-Filter Model and Beta-Divergence Non-Negative Matrix Factorization
A musical source separation algorithm for mono channel signals is presented in this paper. The algorithm is based on a non-negative matrix factorization (NMF) method which factorizes the magnitude spectrum of the input signal into a sum of components, each of which has a fixed magnitude spectra and a time-varying gain. In order to factorize the input spectrum, the input signal is modeled using a source-filter model. The parameters of the source-filter model are estimated by minimizing the beta-divergence from the input spectrum to the reconstructed model. This source-filter model takes advantage of the reliability of the estimated parameter. Simulation experiments were carried out using mixed signals composed of piano and cello. The performance of the proposed algorithm was compared to the basic NMF algorithm using a linear signal model and to a source-filter model NMF algorithm using Kullback-Leibler divergence instead of beta-divergence. According to the results of these simulations, the proposed algorithm has a better separation quality than that found in the previous algorithms.
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member and would like to subscribe to the E-Library then Join the AES!
This paper costs $33 for non-members and is temporarily free for AES members.