Single-channel speaker-dependent speech enhancement exploiting generic noise model learned by non-negative matrix factorization

This paper considers the single-channel speech separation problem given a noisy observation recorded by a microphone. More precisely, we focus on the speaker-dependent approach where spectral characteristic of target speech is learned in advance from a clean example. In training process, we propose to learn a generic spectral model for noise source by collecting various types of environmental noise via the established non-negative matrix factorization framework. In speech enhancement process, we propose to combine two existing group sparsity-inducing penalties in the optimization function and derive the corresponding algorithm for parameter estimation based on multiplicative update (MU) rule. Experiment result over mixtures containing different real-world noises confirms the effectiveness of our approach.