Multimicrophone noise reduction using recursive GSVD-based optimal filtering with ANC postprocessing stage

Recently, a generalized singular value decomposition (GSVD)-based optimal filtering technique has been proposed for enhancing multimicrophone speech signals degraded by additive colored noise. The GSVD-based optimal filtering technique has a better noise reduction performance than standard beamforming techniques provided that the used filter length is large enough. In this paper, it is shown that the same noise reduction performance can be obtained with shorter filter lengths at a lower computational complexity by incorporating the GSVD-based optimal filtering technique in a generalized sidelobe canceller type structure, i.e., by adding an adaptive noise cancellation (ANC) postprocessing stage. Even when using short filter lengths, the total computational complexity is essentially determined by the calculation of the GSVD of a speech and a noise data matrix. It is shown that the complexity can be significantly reduced by using recursive GSVD-updating algorithms and by using subsampling. Simulations have been performed for various acoustic scenarios (different and multiple noise sources and different reverberation conditions), where both the improvement in signal-to-noise ratio and speech distortion have been analyzed. These simulations show that the GSVD-based optimal filtering technique with an ANC postprocessing stage has a better noise reduction performance than standard fixed and adaptive beamforming techniques while introducing an acceptable amount of speech distortion.