As distributing the whole database as part of a synthesis voice may be prohibitively large, especially if multiple voices are required, appropriate pruning of units can be done to reduce the size of the database. This has two effects. The first is to remove spurious atypical units which may have been caused by mislabelling or poor articulation in the original recording. The second is to remove those units which are so common that there is no significant distinction between candidates. Given this clustering algorithm it is easy (and worthwhile) to achieve the first by removing the units from a cluster that are furthest from its center. Results of some experiments on pruning are shown below.
The second type of pruning, removing overly common units, is a little harder as it requires looking at the distribution of the distances within clusters for a unit type to find what can be determined as, ``close enough.'' Again this involves removal of those units furthest from the cluster center, though this is best done before the final splits in the tree, and only for the most common unit types.
As with all the measures and parameters there is a trade off between synthesis resources (size of database and time to select) verses quality, but it seems that pruning 20% of units makes no significant difference (and may even improve the results) while up to 50% may be removed without seriously degrading the quality. (Similar figures were also found in the work described in [7].)