Um Sistema de Conversão Texto-Fala Para o Português Falado no Brasil

Abstract

Text-to-speech synthesis is the subject treated in this work. Most of the difficulties related to this task are discussed, and an implementation of a Brazilian Portuguese text-to-speech concatenative synthesis system is presented. The system uses an inventory of 2,450 pre-recorded speech segments, and is able to employ both TD~PSOLA and hybrid synthesis techniques. The use of carefully chosen linguistic criteria, mainly during phonetic transcription and also during the creation of the speech segments inventory, is the main contribution of this work. The phonetic notation employed distinguishes two kinds of phonetic segments (full and reduced), on the basis of the extension of coarticulation phenomena. The main criterion underlying the building of the speech segments inventory was to preserve reduced segments and vowel clusters. Nevertheless, some of the reduced segments were split, aiming at reducing the size of the inventory. Once again, in this case, specific linguistic criteria were employed, in order to minimize spectral discontinuities after concatenation.