Beta Divergence for Clustering in Monaural Blind Source Separation
General purpose audio blind source separation algorithms have to deal with a large dynamic range for the different sources to be separated. In our algorithm the mixture is separated into single notes. These notes are clustered to construct the melodies played by the active sources. The non-negative matrix factorization (NMF) leads to good results in clustering the notes according to spectral features. The cost function for the NMF is controlled by a parameter beta. The choice of beta depends on the dynamic difference of the sources. The novelty of this paper is to propose a simple classifier to adjust the parameter beta to current dynamic ranges for increasing the separation quality.
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.