Theoretical analysis of musical noise generation for blind speech extraction with generalized MMSE short-time spectral amplitude estimator