Music Enhancement Using Nonnegative Matrix Factorization with Penalty Masking

The implementation of nonnegative matrix factorization may simply be carried out on a stepwise basis. However, this mechanism could lead to longer convergence time than that by designing a suitable cost function with additional penalty terms. Moreover, the decomposed matrices do not necessarily reflect the real situation and physical meaning the block-oriented decomposition is processing. For music enhancement in the noisy environment, this shortcoming may be alleviated by imposing strong constraints on noise components via iteratively adaptive processing. To characterize the ever-changing features in their frequency with respect to time, we also introduce a perception based filtering as the preprocess for more reliable noise detection. As a result, one example of numerical design based on a reverberant room recording is shown to demonstrate the usefulness of our approach.