Low distortion speech denoising using an adaptive parametric Wiener filter

This paper describes a parametric Wiener filter designed for noise removal with low distortion of the speech signal. The classic Wiener filter is augmented with a proportional variable for noise estimation, and a floating floor variable for the transfer function. These two variables are adaptive to the estimated noise energy in parametric relations, determined experimentally for the corresponding noise estimator. The optimization of those parameters can enable the filter to achieve low distortion noise removal. Experiments using some office and home appliance noises have shown superior performance in comparison to the common Wiener filter and the spectral subtraction approaches. The proposed method has comparable quality but less computational demands than the psychoacoustically motivated Gustafsson filter. Because of low distortions, the filter may also be used in cascade with others to achieve better total performance.