Recording script design for a Brazilian Portuguese tts system aiming at a higher phonetic and prosodic variability

Abstract

The naturalness of intonation of the speech produced by corpus-based synthesis systems is strongly dependent on factors such as prosodic modeling algorithms, quality of the speech corpus, speakerpsilas voice characteristics, and recording script. In order to improve naturalness, the recording script must present a considerable amount of variants of a given phone in distinct prosodic contexts, i.e., this corpus must own an appropriate phonetic as well as prosodic variability. Aiming to improve the variability of such a corpus, this work proposes a procedure to design the recording script for a concatenative speech synthesis (text-to-speech - TTS) system developed for the Brazilian Portuguese (BP) language. In this design procedure, four stages are considered: grapheme-to-phoneme conversion, prosodic annotation, feature vector representation, and selection itself. The procedure of prosodic annotation is an original contribution of the current research work. The selection itself is carried out by using an approach based on genetic algorithms. In addition, a set of 3000 prosodically rich BP sentences will be made available as a research tool for speech processing applications.