A subspace approach for enhancing speech corrupted by colored noise