Spectral Masking and Filtering
Timo Gerkmann, Emmanuel Vincent · 2018
In this chapter, we consider spectral masking filters for interference reduction in case of a single-channel input. The considered techniques are thus relevant when only one microphone is present or to post-process the output of a beamformer. After discussing the basic concept of spectral masking, we review different methods for determining spectral masks given the signal statistics. We start with simple spectral subtraction rules and linear Wiener filtering. We then introduce the powerful tool of Bayesian estimation and argue that this technique is advantageous especially when estimating nongaussian quantities such as spectral magnitudes. Finally, we address ways to improve the perceptual quality of the enhanced signals by using heuristic tweaks, by incorporating a compressive function in the estimation, and by sophisticated smoothing methods that aim at reducing spectral outliers while keeping target distortion low.