NMF-based speech enhancement using multitaper spectrum estimation

In the nonnegative matrix factorization (NMF) method, the clean speech dictionary, noise dictionary, and the activation matrix represent the key elements that greatly affect the speech enhancement performance. Current implementation of this framework employs the magnitude (or the power) spectrum coefficients computed from the short-time Fourier transform (STFT) as input features, which are characterized by an increased estimator variance. In this paper, we investigate the effects of using multitaper spectral estimation in the learning of the NMF speech/noise dictionary and activation matrix, on the quality of the Wiener filter used to predict the clean speech spectrum. The evaluation of the proposed method for various types of noise and input SNR show substantial speech quality improvements in terms of PESQ, SDR and SSNR measures compared to the NMF system based on conventional windowed periodogram estimation. We also find that the noise dictionary and the activation matrix of the noisy speech are the most important elements that benefit from the multitapering approach in the NMF system.