Stationary noise suppression is widely used, mostly for reducing noise in speech signals or for audio restoration. Most noise suppression algorithms are based on spectral modification, i.e. a real-valued gain filter is applied to short-time spectra of the speech signal to reduce noise. The more noise is to be removed, the more likely are artifacts due to aliasing effects and time variance of the gain filter. A perceptually motivated systematic time and frequency smoothing of the gain filter is proposed to improve quality, considering the frequency resolution of the auditory system and masking. Comparison with a number of previous methods indicates that the proposed noise suppressor performs as good as the best other method, while computational complexity is much lower.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.