Rank reduction is a common noise-reduction technique in signal processing. We analyze a class of rank-reduction algorithms based on orthogonal projection on certain subspaces, and show that the properties of these algorithms can be compared by means of FIR filters defined by the canonical vectors associated with the projections. We use our new analysis to demonstrate that ULV decompositions work well in connection with speech signals, also in the absence of a gap in the singular values (which is usually assumed to be present).