Speech unit selection based on target values driven by speech data in concatenative speech synthesis

In concatenative speech synthesis systems, speech features (prosody, etc.) are estimated as "target values" according to the input text at the start of processing. The target values are used to evaluate each speech unit in the speech database of the system in order to find out a "nearest" unit, which shows the lowest cost, to the value. In conventional systems, the target value is treated as an absolute value to calculate the cost of each unit. In this paper, we propose a new method to drive the target value depending on the features of speech units in the synthesis system within the range in which naturalness and clarity are maintained. This method was applied to fundamental frequency control, which is one of the most important factors in maintaining naturalness of synthesized speech, based on a superpositional model for a 10-name (Japanese place names) synthesis test, and it was found that the cost decreased an average of 22.9%.