Modelling Japanese intonation using PENTAtrainer2

Abstract

This paper presents results from Japanese intonation modelling using PENTAtrainer2, an articulatory synthesiser. Our first aim is to show that PENTA, on which PENTAtrainer2 is based, can achieve high accuracy in predictive synthesis of varying intonation contours. We trained the synthesiser on a 6251-sentence functionally annotated corpus and generated F0 contours for each communicative condition. The accuracy of speaker-dependent and independent synthesis, together with naturalness ratings, show that PENTA is effective in modelling Japanese intonation. This suggests that once contextual variability is incorporated into a model, multi-functional targets alone would suffice as the prosodic representation even in a sizeable corpus.