The HMM implementation used for these experiments does not include null states, which have been used by other researchers for modeling deletions. To approximate the effect of null states, we can add skip edges around each state. That is, for each path of three states (a-->b-->c) in the HMM, we can add a new edge around the middle state (a-->c).
In some models, one state will have many in- and out-edges, and the number of skip edges around that state would be enormous. Since such nodes are usually junk loops, skip edges around them generally will not save many bits. To avoid the potential quadratic expansion of the number of edges, skip edges around a state are not added for states which would produce too many skip edges. The threshold is currently set at about one fifth the number of states.
The skip edges need to be given a fairly high cost, to avoid stealing
too much probability from the other out-edges of a. For example, the
hmm700b script described in Section 3.1 
adds skip
edges with a cost of 6 bits (probability approximately 0.016) at the end of
the script.