A novel Bayesian framework for speech enhancement using speech presence uncertainty

In this paper, we propose a codebook-based speech enhancement scheme under a novel Bayesian framework. The goal of this framework is to optimally combine power spectral envelopes (PSEs) estimation of speech and noise from speech absence hypothesis and speech presence hypothesis in current frame. Then, the noisy observation is passed through the reconstructed Wiener filter to obtain the enhanced amplitude spectrum. In order to further improve the speech quality, we exploit one geometric equation to derive the phase spectrum of the clean speech. Finally, both enhanced amplitude and estimated phase spectra are used for recovering the clean speech signal. Experiments show that the proposed approach generates considerable improvement compared to the conventional codebook-based methods.