In this paper a new approach is presented to develop the subspace-based speech enhancement for non-stationary noise cases. The new method updates the noise correlation matrix segment-by-segment assuming that only the eigenvalues of the matrix are varying with time. In other words, the characteristic of varying loudness of noise signals is just considered, as it is observed in the modulated white noise case where the eigenvectors are invariant over time. The proposed scheme for updating noise correlation matrix is embedded in the framework of a soft model order based subspace approach for speech enhancement. The experiments show significant improvement in different non-stationary noise types.