Study Of Microphone Array Beamforming For Speech Communication

Microphone arrays can be used to solve many important acoustic problems such as distant sound acquisition,speech enhancement,noise and interference reduction,dereverberation,spatial sound acquisition,sound source localization and tracking,etc.Consequently,they have been widely used in a wide range of applications such as hands-free speech communication,robotics,teleconferencing,telecollaboration,human-machine interfaces,to name but a few.A core component of microphone array system is the so-called beamforming,which is to estimate a signal of interest while suppressing noise,interference,and reverberation by exploiting the spatio-temporal information embedded in the signals observed by the array sensors.While it has been intensively studied for several decades in as diverse fields as radar,sonar,wireless communications,and seismology,beamforming to deal with acoustic signals is a non-trivial task due to complicated reasons including but not limited to the following.1)Speech and audio signals are broadband in nature and their frequencies can range from a few tens of Hz to 20 kHz.It is difficult if not impossible to design effective beamformers that can perform uniformly over such a large frequency range.2)The number of sensors in a microphone array is often limited by practical factors such as the cost and size of the devices where the array is used.It is challenging to achieve satisfactory array gains with a small aperture and a small number of microphone sensors.3)Non-stationary noise generally exists in the applications where microphone arrays are equipped,and there also exists sensor self noise and mismatch between different sensors;the adverse noise environment makes it difficult to achieve robust and large array gain,at the same time.Nevertheless,a great deal of efforts has been devoted to beamforming with microphone arrays in the literature.Broadly,four different principles have demonstrated potential for high-fidelity sound acquisition,leading to four different yet connected beamforming methods,i.e.,differential beamforming,superdirective beamforming,frequencyinvariant beamforming,and adaptive beamforming.The major theoretical contributions of this dissertation are summarized as follows.1.A multistage-cascaded beamforming framework is proposed,and the spatial Z transform is subsequently introduced to analyze and investigate the framework.Then,differential beamforming of different orders is systematically studied,and a reduced-order robust differential beamforming method is proposed to overcome the white noise amplification problem of differential beamforming.2.A robust superdirective beamforming based on a two-stage-cascaded framework is proposed.The proposed approach can improve the robustness of the traditional su-perdirective beamformers and,at the same time,maintain the frequency invariance of their beampatterns.3.Based on the Jacobi orthogonal polynomial expansion,a principle measuring the frequency invariance of a beampattern is built.With the proposed principle,the optimal beamformers cannot only control the beampattern degradation to form frequency-invariant beampattern,but also control the error distribution between the designed and the desired beampatterns.4.By investigating the performance of the widely-known MVDR beamformer as a function of source incidence angle in different noise environments and reverberation conditions,it is revealed that,for linear arrays,the best performance is generally obtained when the desired source is at their endfire directions.This conclusion can provide guidance for designing better array systems in practical applications.