Study Of Microphone Array Speech Enhancement For Speech Communication And Human-machine Interfaces

In acoustic applications such as voice communication and human-machine speech interfaces,the acoustic signal of interest picked up by microphone sensors is inevitably contaminated by unwanted noise,which will impair not only speech quality but also speech intelligibility.Speech enhancement,which attempts to recover the signal of interest from noisy observations using signal processing techniques,has been widely studied.Among all the methods developed in the literature,two commonly used methods are beamforming and mutlichannel noise reduction,both are based on the use of microphone arrays.Since speech and audio signals are broadband in nature,whose frequencies typically span between 20 Hz and 20 kHz,how to enhance such signals without introducing much signal distortion has been a great challenge for both beamforming and multichannel noise reduction.The work of this dissertation is therefore organized to address this challenge with its focal point being placed at frequency-invariant beamforming with circular microphone arrays(CMAs)and concentric CMAs(CCMAs),Krylov superdirective beamforming,and multichannel maximum signal-tonoise ratio(SNR)filters for noise reduction.The major contributions of this dissertation are as follows.?It shows that the optimal approximation of the beamformer's beampattern from a least-squares error(LSE)perspective is the Jacobi-Anger expansion.Based on the Jacobi-Anger expansion,an approach is developed to design frequency-invariant beamformers with CMAs(FIB-CMAs),whose main beam can be steered to any directions in the sensor plane.With the proposed FIB-CMAs,an explicit form of the white noise gain(WNG)and the directivity factor(DF)is deduced,which explains clearly the effect of the number of microphones and the array radius on beamforming performance.It is also shown that the circular differential microphone arrays(CDMAs)and circular harmonics beamformers(CHBs)are particular cases of the developed method.?Through theoretical analysis,it is shown that the nulls problem with FIB-CMAs is caused by zeros of the Bessel functions in the denominators of the beamformer's coefficients.To mitigate this problem,a frequency-invariant beamforming method with CCMAs(FIB-CCMAs)is developed,which combines different rings of microphone sensors together with appropriate radii.This method does not require the microphone sensors in different rings of the CCMA to be aligned,which is useful in practice,particularly when microphone arrays with small and compact apertures have to be used.On the design of robust FIB-CCMAs,we propose to approximate the beamformer's beampattern corresponding to a different ring with the Jacobi-Anger expansion of a different order,depending on the number of sensors.It is shown that one can obtain an Nth-order desired and symmetric beampattern as long as one ring has 2N+1 or more sensors.With this method,the array structure can be very flexible since the inner rings can have fewer microphones than the outer rings and the microphones in different rings do not need to be aligned.Also discussed are the conditions for the design of FIB-CCMAs with commonly used array geometries and the advantage of adding one microphone at the center of either a CMA or a CCMA,i.e.,circumventing the deep nulls problem caused by the 0th-order Bessel function.?An approach to superdirective beamforming is developed based on the use of the socalled Krylov matrix.By properly choosing the dimension of the Krylov subspace,the developed beamformers can make a good compromise between high DF and a reasonable level of white noise amplification.This principle is also extended to the design of the Krylov maximum front-to-back ratio,parametric superdirective,and parametric supercardioid beamformers.?A multichannel maximum SNR filter is developed for noise reduction in the shorttime Fourier transform(STFT)domain,which takes into consideration of both the interchannel and interframe information.The maximum SNR filter is formulated by maximizing the output SNR with the minimum distortion constraint,which can significantly increase both the SNR and the speech quality.