Musical Noise Reduction Method Based on MMSE Combined with Wiener Filtering.

The purpose of this paper is to reduce the distortion caused by musical noise and inaccurate phase estimation of pure signal. Firstly, the Wiener filter and the minimum square error (MMSE) estimation in series connection to construct a two-stage hybrid system (W-MMSE) in frequency domain to obtain superior gain function. Then, the phase spectrum of the noisy speech signal is reconstructed, and the expected signal is recovered by combining the pure speech amplitude spectrum estimated by means of the system with the phase spectrum of the reconstructed noisy signal. In order to evaluate the performance of the proposed method, the NOIZEUS dataset is used for simulation. The objective experiments of perceptual estimation of speech quality (PESQ) and log likelihood ratio (LLR) have proved that W-MMSE can substantially suppress the noise signal, eliminate the musical noise, and improve the speech quality and intelligibility.