Nonnegative Matrix Factorization with Adaptive Elements for Monaural Audio Source Separation

For monaural audio source separation, Nonnegative Matrix Factorization (NMF) has become one of the most dominant methods. This thesis contributes adaptive methods in the context of two extensions: Nonnegative Matrix Factor Deconvolution (NMFD) and constraints for NMF computation. It is shown that these elements improve separation results only under specific conditions. Consequently, adaptive algorithms are developed in this thesis. The underlying hypothesis of these modifications is that NMF components have different properties depending on the acoustical event to which they correspond.
In the context of NMFD, a generalization of NMFD is proposed which makes it possible to use the deconvolution approach only on a subset of the components. Further, with this approach it is possible to adapt the parameters of NMFD on each component individually. Experimental results show that an algorithm using the generalized NMFD leads to better separation results as a comparable algorithm using NMF or NMFD.
For NMF with additional constraints, two adaptive extensions are presented. The first extension adapts the constraints depending on the properties of different NMF components. As a result, these constraints are imposed stronger on the components for which they are beneficial and weaker on others. Secondly, an algorithm is developed, which also makes it possible to adapt the strength of the constraints to different entire mixtures during runtime of the NMF. Experimental results show that both algorithms are beneficial for the source separation results.
The proposed adaptive elements for NMF prove to be an effective addition to the state of the art of NMF, enabling improved quality of fully automatic blind source separation of monaural audio mixtures.