| In speech communication systems,the speech signal of interest picked up by microphone is inevitably contaminated by unwanted noise,which can cause dramatic degradation in speech quality and intelligibility.Therefore,speech enhancement techniques are required to recover the “clean” speech from its noisy observations.This problem has been intensively studied since 1960 s when the first speech enhancement method was invented.Numerous speech enhancement methods have been developed over the last six decades.Although significant progress has been made,the existing speech enhancement methods are still far from being sufficient in dealing with noise and interference in practical environments.This dissertation is therefore devoted to address challenges and develop new beamforming and multichannel noise reduction methods for speech enhancement.The major contributions are as follows.(1)By investigating and reorganizing the signal model of the time-domain multichannel noise reduction,a bilinear noise reduction scheme is proposed,which separates the temporal and spatial processing of the noise reduction filter.As a result,instead of computing a long-length noise reduction filter,one can compute two sub-filters of much shorter length,which leads to two advantages over the traditional methods:one is the significantly lower computational complexity,and the other is that much less data samples are needed in computing the shorter sub-filters,which enables a better tracking of spectral and temporal variations of noise signals.(2)Based on the singular value decomposition(SVD)and Kronecker product decomposition(KPD),we investigate a low-rank representation of the traditional spatiotemporal noise reduction filters and develop a KPD based noise reduction framework,which decomposes the traditional multichannel noise reduction filter into two subfilters of much shorter length,one for the temporal dimension and the other for the spatial dimension.The bilinear noise reduction scheme is shown to be a particular case of this KPD based noise reduction framework.(3)For processing broadband and scattered acoustic sources with small-spacing microphone arrays,we develop a maximum diffuse noise gain(MDNG)beamformer using a joint diagonalization technique,which is effective in suppressing diffuse noise,but at the price of low white noise gain(WNG).We also propose a maximum WNG(MWNG)beamformer,which is robust to array’s imperfections,but paying a price of sacrificing the diffuse noise gain(DNG).To make a tradeoff between the WNG and DNG,we propose a raised-rank MDNG(RRMDNG)beamformer,which on the one hand,can achieve high DNG and,on the other hand,is robust to implement.The MDNG and MWNG beamformers are particular cases of this general method.(4)To deal with scattered interference incidents from a pre-specified range of directions of arrival with small-spacing microphone arrays,we develop an approach to the design of beamformers by incorporating spareness constraints.The design process is formulated as a constrained LASSO problem.By adjusting a tunable parameter,the resulting beamformer can be made robust to sensors’ self noise and array’s imperfections while achieving high directivity factor(DF)as well as high signalto-interference ratio(SIR)gain to attenuate scattered interference incidents from a given a priori angle range.(5)In many application scenarios,interference and reverberation may mainly come from a certain region,and it is therefore necessary to develop beamformers that can preserve signals of interest while minimizing the power of signals coming from the region where interference and reverberation dominate.For this purpose,we first reexamine the so-called front-to-back ratio(FBR)and the related classical supercardioid beamformer,and then extend the FBR to a generalized FBR(GFBR)and investigate the relation between the FBR/GFBR and DF.Based on these,a set of raised-rank and regularized beamformers are deduced by using a well-known joint diagonalization technique,which can make compromises between the FBR/GFBR and WNG or DF flexibly for dealing with the white noise amplification problem and the limited DF associated with the supercardioid beamformer. |
