Automatic stylization, coding and modelling of intonation in text-to-speech for under-resourced languages

Abstract

In this paper an automatic method to implicitly model intonation for statistical parametric speech synthesis (SPSS) is presented. The approach is ideally suited to single speaker speech databases as used in text-to-speech (TTS), due to the models being speaker-specific. Fundamental frequency curves are automatically stylized based on the speaker-specific acoustics in the recorded database, requiring no models rooted in linguistic theory, and therefore being well suited to intonation modelling in under-resourced languages. The stylized curves are then coded into abstract pitch labels, which are used as features in the training of the statistical parametric acoustic models. A conditional random field (CRF) model is trained in order to predict the abstract pitch labels from the text for synthesis. The CRF model can be used to predict the abstract pitch labels on the syllable, word and phrase tiers. Objective and subjective results on synthetic voices built from English and isiXhosa speech databases are shown.