An improved generalized weighted Bayesian estimator for speech enhancement

In this paper, an improved Bayesian estimator for single-channel speech enhancement is studied. Based on the observation that the speech discrete Fourier transform (DFT) coefficients are better modeled by a Chi distribution and the human auditory system properties, a generalized weighted Bayesian estimator with Chi priori is proposed. Then, with the speech presence uncertainty (SPU) estimation, a novel Bayesian estimator incorporating the SPU with Chi priori is derived. Moreover, a hybrid priori SNR (HSNR) estimator is also studied which can eliminate musical noise effectively in the speech onset and offset periods by employing voice activity detection (VAD) and temporal cepstrum smoothing (TCS) techniques. Simulation results show that a significant improvement is achieved by the proposed algorithm.