Subband spectral-subtraction speech enhancement based on the DFT modulated filter banks
Yu Cai, Chaohuan Hou · 2012
In this paper, we propose a new algorithm for suppressing the additive background noise, especially colored, in speech signals. An oversampled DFT modulated filter bank is designed to decompose the time series into equal spaced subbands. In each subband, an improved spectral-subtraction technique is applied. This method mainly considers that most environmental noises affect the speech spectrum non-uniformly within different frequencies and multiband processing will be more accurate and effective. Compared to both the conventional and multiband spectral-subtraction approaches, objective evaluation reveals that this algorithm performs better in speech quality improvement. Informal listening tests also show a reduction of musical noise.