The Research Of Speech Enhancement Methods Based On Statistical Model And Improved Phase Spectrum

For a long time,speech enhancement is used to handle speech signal degraded by kinds of interference.The goal is to recover the original clean speech as much as possible and improve the speech quality and intelligibility.Depending on the number of audio channels,speech enhancement can be divided into two categories:single channel speech enhancement and multi-channel speech enhancement.While the single channel speech enhancement method is more common than the multi-channel one,so this paper mainly focuses on the methods of single channel speech enhancement.Among a very large number of single channel speech enhancement approaches,the speech enhancement methods based on the statistical model is one of the most popular technologies because of their large range of applicable signal-to-noise ratio(SNR)and wide applicable noise environment.This research analyzes two speech enhancement methods based on the Gaussian distribution model,i.e.,the minimum mean-square-error spectral power(MMSE-SP)and the minimum mean-square-error magnitude-squared spectrum(MMSE-MSS).After expounding the respective advantage and disadvantage,we propose a new method by combining the speech absence probability(SAP),namely the SPMSS-SAP.The experimental results indicate that the proposed method in our study has a better performance of removing background noise and improving the speech quality.In addition,most of the existing speech enhancement methods based on Fourier transform only focus on the magnitude spectrum estimation or its modification,while the effect of phase spectrum is ignored.In previous works,many scholars reported that the noisy phase spectrum is the best estimation of real phase spectrum,and some people even considered that the phase spectrum hardly carries any useful information.But recent studies found that the phase spectrum plays an important role in speech intelligibility and accurate phase spectrum estimation that would significantly improve speech quality and intelligibility.The importance of the phase spectrum gradually is attracted the attention of the public,and some achievements have be obtained.The phase spectrum compensation(PSC)method is the product of this idea,which is based on the modification of the phase spectrum modification.This paper discusses the PSC method in detail and proposes a new method based on phase spectrum modification which is called modified PSC(MPSC).Adequate experimental results show that the proposed MPSC method significantly outperforms the original PSC method.