Spectral domain speech enhancement method based on noise compensations in both magnitude and phase spectra

Abstract

In this thesis, a noisy speech enhancement method based on noise compensation performed on short time magnitude as well phase spectra is presented unlike the conventional spectral subtraction method. Here, the noise estimate to be subtracted from the noisy speech spectrum is proposed to be determined exploiting the low frequency regions of noisy speech of current frame rather than depending only on the initial silence frames. We argue that this approach of noise estimation o ers the capability of tracking the time variation of the non-stationary noise thus resulting in a noise compensated magnitude spectrum. By employing the noise estimates thus obtained, a procedure is formulated to compensate the distortion in the phase spectrum,which is kept unchanged in the typical speech enhancement methods. The noise compensated phase spectrum is then recombined with the noise compensated magnitude spectrum to produce a modi ed complex spectrum thus synthesizing an enhanced frame. Extensive simulations are carried out using NOIZEUS database in order evaluate the performance of the proposed method. It is shown in terms of objective measures, spectrogram analysis and informal subjective listening test that the proposed method consistently outperforms some of the state-of-the-art methods of speech enhancement from noisy speech corrupted by white or train or babble noise of very low levels of SNR.