Robustness of Svd-based Optimal Filtering for Noise Reduction in Multi-microphone Speech Signals

This paper discusses an SVD-based signal enhancement procedure, applied to noise reduction in multi-microphone speech signals. The SVD-based signal enhancement procedure amounts to a specific optimal filtering problem when the so-called ‘desired response’ signal cannot be observed. The optimal filter can then be written as a function of the generalized singular vectors and singular values of a speech and noise data matrix. It is shown that the SNR improvement provided by the SVD-based optimal filtering technique is better than the improvement obtained with standard beamforming techniques. Moreover most beamforming techniques assume the position of the speech source and the microphone array configuration to be known. Therefore the performance of these techniques is rather sensitive to deviations from these assumptions, i.e. incorrect estimation of the position of the speech source and uncalibrated microphone arrays. It is shown that the SVD-based optimal filtering technique is more robust to such deviations than standard beamforming techniques.