A new approach for speech enhancement based on a constrained Nonnegative Matrix Factorization

In this paper, a new approach is presented for single-channel speech enhancement which is based on Nonnegative Matrix Factorization (NMF). The proposed scheme combines the noise Power Spectral Density (PSD) estimation based on a constrained NMF and Wiener filtering to enhance the noisy speech. The imposed constraint is motivated by the time correlation of the underlying observations and enforces the NMF to give smoother estimates of the nonnegative factors. Compared to the standard NMF approach and Wiener filtering based on a recently developed noise PSD estimator, Source to Distortion Ratio (SDR) is improved for the evaluated noise types for different input SNRs.