Robust equalization of multichannel acoustic systems

• Main Author: Zhang, Wancheng
• Other Authors: Naylor, Patrick
• Published: Imperial College London 2010
• Subjects: 621.382
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.519297

In most real-world acoustical scenarios, speech signals captured by distant microphones from a source are reverberated due to multipath propagation, and the reverberation may impair speech intelligibility. Speech dereverberation can be achieved by equalizing the channels from the source to microphones. Equalization systems can be computed using estimates of multichannel acoustic impulse responses. However, the estimates obtained from system identification always include errors; the fact that an equalization system is able to equalize the estimated multichannel acoustic system does not mean that it is able to equalize the true system. The objective of this thesis is to propose and investigate robust equalization methods for multichannel acoustic systems in the presence of system identification errors. Equalization systems can be computed using the multiple-input/output inverse theorem or multichannel least-squares method. However, equalization systems obtained from these methods are very sensitive to system identification errors. A study of the multichannel least-squares method with respect to two classes of characteristic channel zeros is conducted. Accordingly, a relaxed multichannel least- squares method is proposed. Channel shortening in connection with the multiple- input/output inverse theorem and the relaxed multichannel least-squares method is discussed. Two algorithms taking into account the system identification errors are developed. Firstly, an optimally-stopped weighted conjugate gradient algorithm is proposed. A conjugate gradient iterative method is employed to compute the equalization system. The iteration process is stopped optimally with respect to system identification errors. Secondly, a system-identification-error-robust equalization method exploring the use of error models is presented, which incorporates system identification error models in the weighted multichannel least-squares formulation.