Unit selection systems typically select from a finite set of units in the database. They are looking for the best path through a given set of units. Of course when there are no examples of good units in that set, this can viewed either as a lacking in the database coverage or that the desired sentence to be synthesized is not in domain.
Many system do some localised smoothing at boundaries. While [12] introduces the notion of fusion units. Thus he effectively increases the number of units available for selection by allowing the construction (fusion) of new units from the existing ones. This direction will gives us a more general solution towards a more general set of units.
A even more general solution is that taken by HTS, [13]. Using a HMM-based framework, in contract to [9] which selects sub-parts of the database, HTS uses the HMM parameter representation to generate the speech. Thus effectively a much wider range of units is available, as context affect generation through constraining deltas, and smoother joins are possible. There is a cost though. In its basic form the excitation part of the signal is not modelled thus reducing the quality to vocoded speech, though better excitation modelling is being worked on.
What is important about such directions in unit selection is that the size of the inventory is effectively much larger. However, although it can potentially cover the given space better than conventional unit selection synthesis systems, it its still limited by the examples in the database.