Japanese intonation synthesis using superposition and linear alignment models

This paper outlines a new approach to Tokyo Japanese intonation synthesis, in which the F0 contour of an utterance is generated using the superposition of multi-level phrase curves and lexical accent curves, coupled with linear alignment models which determine the precise alignment of the curves with the segmental material. We first discuss the construction of a phrase curve used to model the prosodic domain termed the 'UA-group' (defined below), and describe the alignment of this curve with the syllabic structure of the utterance. Then, we describe a separate accent curve, carrying independent prominence specification, which is added to this UA-curve in the case of accented phrases. The alignment of the accent curve with the segments is determined by linear alignment models.