Abstract:

Conventional postfiltering (CPF) algorithms often use a fixed filter bandwidth to estimate the auto-spectra and the cross-spectrum. The drawback of the CPF algorithm under the stochastic model and the ways to improve the performance of the CPF algorithm are studied firstly. To improve noise reduction without introducing audible speech distortion, a novel spectral estimation method based on the structure of the noise power spectral density (NPSD) is proposed. The proposed spectral estimation method is used to improve the performance of the CPF. Experimental results verify that the proposed algorithm is better than the CPF algorithms in terms of the segmental signaltonoise ratio improvement and the noise reduction, especially the noise reduction is about 6 dB higher than the CPF.