/m152: You are right, there are no thresholds in the Viterbi algorithm.

Generally you might find some way of tuning the training and/or prediction algorithm to produce overall more or fewer hyphens. But I can't say I know a good way to do this.

More generally, per-letter accuracy, per-word accuracy, and log-likelihood are all different objective functions. A research question is how to tune CRF-type methods in order to maximize (on test data!) whatever objectove makes most sense for the application.