On sentence-level factors governing segmental duration in Japanese

In this paper, the durational characteristics of Japanese are statistically analyzed with the aim of establishing a fine duration setting rule. First, a statistical duration control model is formulated through factor analysis on a large-scale isolated-word database (5240 words). Second, using 282 fluently spoken connected speech utterances, segmental durations calculated by this model are compared with measured durations. Error analysis made clear the existence of the following sentence-level duration effects: (1) Prepausal lengthening is greater than simple word-final lengthening; and (2) shorter durations are found in sentence-final position in read sentences than in the other positions. By incorporating these effects into the duration-setting rules, the prediction error for sentences were reduced to the same level as for isolated words.