Base types and searching

An alternative solution, which we will adopt here, is called unit back-off. In this, we manufacture new units from pieces of existing units. Thus, instead of using a /t-iy/ unit instead of a /t-ih/ unit, we instead use only the first part of /t-iy/, and join this to the second half of a /d-ih/ unit to create the required /t-ih/ diphone. There is no standard way of doing this, but, since the number of missing diphones is often small, it is not too difiicult to specify how to build the missing diphones by hand. As in the example just 

Rgure 16.10 An ergodic half-phone search. Here a-(n) means a half-phone unit of the second half of [a] in the context of a following [n] and n-(a) means a half-phone unit of [n] in the context of a preceding [a]. 

If we have N phonemes in om system, we will have 2N half-phone base types and diphones. Assuming that we have on average 500 units for each diphone base type, we would have on average 250A units for each half-phone base type. Thus, if we consider every unit that matches the specification, we now have 250A units at each time slot in the search for half-phones, compared with 500 for diphones, so if A = 40 this means that we have 10000 units, i.e. 20 times more units at each slot. Given that the join function is calculated over every combination of units at a join, this means that instead of 500 = 250000 join costs to calculate we now have (250= 100000000 joins, 400 times more. These calculations are of course for the same-sized database, which 

We should now see that in effect the diphone formulation with the half-phone backoff and the half-phone formulation with restricted join possibilities are in fact identical in terms of what speech is passed to the search. This equivalence between base types, back-off strategies and search topologies can be shown for all base types that share the same type of join (e.g. this is nue for all those that allow joins in phone middles). This therefore shows that the choice of base type really isn t as fundamental as it seems the choices of which features to use and how to formulate die target and join functions are much more significant in determining final quality. 

Using non-uniform units complicates the search somewhat since now we don t have a lattice of time slots where each time slot aligns one-to-one with the specification. We can, however, still perform a fully optimal search. First we define a grammar , which specifies which units can follow which others (this grammar is no different from the rules which state which diphones or half-phones can be joined). Next we expand each large unit into a number of nodes equivalent to the smallest base type in the system. This then gives a lattice with time slots and nodes as before, which can be searched in the same manner as above. This further demonstrates that the impact of unit choice is neutralised 

---