Um modelo para geração de prosódia de palavras em conversores texto-fala para a língua portuguesa falada no Brasil.

Abstract

This work presents a model for automatic generation of the prosody in a concatenative text-to-speech system for Brazilian Portuguese. The model is based on rules and on stressed syllables i n the words, to determine the intonation contours. For such, an acoustic analysis is accomplished in a corpus of words, containing the most several combinations of phonemes for the syllables, and phonetically balanced sentences, for identification of the behavior of the duration and above all the pitch, along oxytons, paroxytons and proparoxytons words, with up to five syllables. In the first stage of the work, a basic structure of a text-to-speech system is presented and the importance of a prosody model for that kind of system is emphasized. The second stage presents the natural form of speech production and important concepts of linguistics. The t h i rd stage presents the linguistic processing, the prosodic processing and the signal processing levels i n a text-to-speech system. A simpler linguistic processing was opted, contemplating the pre-processing and phonetic transcription levels. The concatenative synthesis was also opted, considered the advantages of simplicity, flexibility and, above all, because the processing of the speech signal is made in the own wave form, maintaining the original characteristics of the signal. The fourth stage presents the acoustic units dictionary and the prosodic model with the obtained results. Syllables and demissylables are used as acoustic units of the dictionary. The relatively large, database of acoustic units, incorporates the most important prosodic information, as well as the articulatory characteristics corresponding to the coarticulation phenomena, in order to obtain a good synthesized speech. The selection of the units is accomplished through a structure of weights attributed to the stressed, pre-stressed and post-stressed syllables, considering the curves of intonation analyzed in the corpus. The model was evaluated through informal listening tests i n words and formal tests i n a corpus of 20 phonetically balanced sentences, wich constitute a representative sample of the universe of all the phonemes and units of the dictionary. The tests with 20 sentences were accomplished with 40 listeners and using the MOS scale (Mean Opinion Score), obtaining a score of