A novel speech enhancement method using power spectra smooth in Wiener filtering

In this paper, we propose a novel speech enhancement method by using power spectra smooth of the speech and noise in Wiener filtering based on the fact that a priori SNR in standard Wiener filtering reflects the power ratio of speech and noise in frequency bins. This power ratio also could be approximated by the smoothed spectra of speech and noise. We estimate the power spectra of noise and speech by means of minima controlled recursive averaging method and spectral-subtractive principle, respectively. Then, the linear prediction analysis is used to smooth power spectra of the speech and noise in frequency domain. Finally, we utilize cross-correlation between the power spectra of the noisy speech and noise to modify gains of the power spectra for further reducing noise in silence and unvoiced segments. The objective test results show that the performance of the proposed method outperforms conventional Wiener Filtering and Codebook-based methods.