Towards an intonation module for a portuguese TTS system

In this paper, a correlation between the linguistic structure of the written text and the real intonation behavior of the read speech in European Portuguese language (EP) is presented. It is our belief that intonation behavior in EP can be strongly predicted from two main coordinates: the syntactic structure of the sentence and its pragmatic communicative function, in one way, combined with the phonological and syntactic nature of the words, in the other way. The purpose of our work is to identify in real speech the main intonation elements, which are relevant to speech naturalness as well as to analyze the factors that determine them. This work addresses the cases of declarative/imperative, interrogative and enumerative phrases. Basic categorizations of the intonation elements, in correlation with the underlying factors are presented. General regularities and correlations as well as the resulting rules, that may be a starting point for practical implementation of an intonation module, are presented and demonstrated, under a Fujisaki’s phonetic/physiological approach. The methodology was based on the observation and modeling of a significant prosodic corpus where different intonation patterns occur in a diversity of text structures. It is our goal to contribute with practical techniques and experience in order to perform a more accurate intonation modeling of Text-to-Speech (TTS) applications, using a rule-based approach.