An application of the warped discrete Fourier transform in the perceptual speech enhancement

An application of the warped discrete Fourier transform (WDFT) in the perceptual speech enhancement is of interest. The WDFT allows nonuniform sampling the z-transform of finite length sequence. We focus on the perceptual warping which allocates frequency samples in good accordance with the Bark scale. The WDFT can replace conventional DFT based analysis/synthesis block of spectral weighting method. In the case of the perceptual warping, there is a problem with signal reconstruction because the WDFT matrix is ill-conditioned. This paper addresses the problem of signal distortions generated in WDFT based synthesis block. Spectral characteristics of the reconstructed signal are analyzed and discussed in the context of the perceptual processing. A new extension of the WDFT intended to cancellation of the synthesis error is presented. The new method is also validated in practical speech enhancement system. The results show that the new algorithm outperforms pure WDFT based system.