Combining non-uniform unit selection with diphone based synthesis

This paper describes the unit selection algorithm of a speech synthesis system, which selects the k-best paths over units from a relational unit database. The algorithm uses words and diphones as basic unit types. It is part of a customisable textto-speech system designed for generating new prompts using a recorded speech corpus, with the option that the user can interactively optimise the results from the unit selection algorithm. This algorithm combines advantages of nonuniform unit selection algorithms and diphone inventory based speech synthesis.