Single channel speech enhancement using principal component analysis and MDL subspace selection

We present in this paper a novel subspace approach for single channel speech enhancement and speech recognition in highly noisy environments. Our algorithm is based on principal component analysis and the optimal subspace selection is provided by a minimum description length criterion. This choice overcomes the limitations encountered with other selection criteria, like the overestimation of the signal{plus{ noise subspace or the need for empirical parameters. We have also extended our subspace algorithm to take into account the case of colored noise. The performance evaluation shows that our method provides a higher noise reduction and a lower signal distortion than existing enhancement methods and that speech recognition in noise is improved. Our algorithm succeeds in extracting the relevant features of speech even in highly noisy conditions without introducing artefacts such as \musical noise".