Evolutionary-Based Design of a Brazilian Portuguese Recording Script for a Concatenative Synthesis System

摘要

Modifications of prosodic parameters in concatenative synthesis systems may lead to a degradation in speech quality, especially when significant pitch changes are accomplished. Aiming to avoid large changes in the speech signal parameters, the speech corpus should present segments with phonetic and prosodic features close to the predicted ones. This condition is more often fulfilled by a speech corpus specially designed to be both phonetic and prosodically rich. The design of this corpus is strongly dependent on the script chosen for recording. For such, a procedure to select the recording script of a TTS system is proposed for the Brazilian Portuguese language. Selected sentences include declarative, exclamatory, and interrogative ones. Phonetic and prosodic information are firstly represented as a set of feature vectors. Next, the amount of distinct feature vectors is used as a fitness value for a genetic-based sentence selection. Experimental results point out a considerable improvement in script variability for speech synthesis applications.