Cepstral Smoothing with Reduced Computational Complexity

Abstract:
It has been recently shown that cepstral smoothing methods may increase the performance of discrete Fourier transform (DFT) based speech enhancement frameworks because spectral outliers can be eliminated while keeping speech distortions low. This can be achieved because speech is very compactly represented in the cepstral domain, enabling a selective smoothing of speech related coefficients and the remaining coefficients. The computational costs of temporal cepstrum smoothing are dominated by two DFTs for the cepstral transform and its inverse. In this work we split the spectrum into D bands and compute D cepstra thereof. Then, for D-1 spectra a pruned discrete cosine transform (DCT) is used which may result in relative computational costs of 37.5% as compared to an unpruned DCT, and 8.3% as compared to an unpruned complex DFT.