Unit-selection-style approaches

The unit selection synthesis technique described in Chapter 16 uses an entirely data driven approach, whereby recorded speech waveforms are cut up, rearranged and concatenated to say new sentences. Given the success of this approach in normal synthesis, a number of researchers have applied these algorithms to FO synthesis [296], [310], [311]. 

FO contours extracted from real utterances are by their very nature perfect. The basic idea therefore is to collect a database of naturally occurring FO contours, and use these at s mthesis time for new utterances. It is extremely unlikely that one of the complete contours in the database will be exactly what we want for this to be the case we would have have to have exactly the same text for the synthesis and database utterance. At a smaller scale, say at the phrase, word or syllable level, it is however possible to find exact or close matches and so the database processed to create a set of FO units. Each unit has a set of features describing it, and at run time, these features are compared with features generated by the prosodic prediction. Sometimes exact matches are found. 

The idea then is to find the best sequence of FO imits for our desired features, and this is calculated by summing the target and join costs for a complete sequence of units. Formally, this is given by  

The advantage of this approach is that we are guaranteed to synthesise a natural contour the only question is whether it will be the correct natural contour for the 

---