NOISE REDUCTION IN MULTI-MICROPHONE SPEECH SIGNALS USING R ECURSIVE AND APPROXIMATE GSVD-BASED OPTIMAL FILTERING

This paper describes some techniques for reducing the computational complexity of a GSVD-based optimal filtering techniq ue for noise reduction in multi-microphone speech signals. It has been shown that this GSVD-based optimal filtering technique has a better noise reduction performance than standard beamforming techniques and is more robust to deviations from the nominal situation [1] [2]. However the computational complexity of this technique is too high to be amenable for real-time implementation. First, the computational complexity is reduced by using recursive and approximate (so called square root-free) GSVD-updating techniques, without a significant loss in performance. Seco ndly, the complexity is reduced by using downsampling techniques. A drawback of using downsampling techniques is slower convergence towards the optimal filter, which is however not a major probl em when considering quite stationary acoustic environments.