SRS prosody rules for Japanese

Abstract

SRS, Hertz's language-independent synthesis rule development system, has been used recently to write a set of phoneme-based rules for Japanese. This paper will describe the Japanese prosody rules, which operate on phonological phrases of various sizes to generate natural-sounding prosodic patterns. The phrases are demarcated in the romanized input string by an easy-to-learn annotation scheme. The annotations are translated into boundary symbols and associated boundary features. These features, together with accent marks, are used to generate a segment-by-segment specification by prosody features, such as [low] (low-pitched) and [ace] (accented). Synthesizer parameter rules then refer either to the boundaries' features or to the intervening segments' features to produce various aspects of an utterances' prosodic pattern, such as phrase-final lengthening or the post-accentual pitch fall. Thus, although SRS in its present implementation allows only one level of explicit user-defined synthesis unit, it is relatively easy to set up implicit units of the larger sizes needed to write straightforward prosody rules.