Consistent multichannel signal separation with wavelet-transformation and modified polyphase structure for spectral subtraction of speech signal

The spectral subtraction method is used to improve the speech signal (SS) quality since this method can essentially reduce or even eliminate SS additive noises. On the contrary to the earlier used transformation with a uniform channel width, two variants of structure are considered, in which the total signal frequency band is subdivided into bands with increasing widths. Such a method of the SS total band subdivision is more consistent with the virtual filtering system of the human auditory system, namely, the so-called perceptual model that manifests peculiar features of perception of SS and other acoustic signals. The paper considers two structures: one based on the discrete wavelet-transformation, the other being a nonuniform system of consistent multichannel separation based on the polyphase structure.