A scheme of syllable duration prediction and F0-contour generation to synthesize Chinese speech

20125 syllable-timing data have been investigated to get their mean syllable durations and their initial/final timing structure. To calculate the actual duration of a syllable with different tones based on its mean duration, tone coefficient /spl lambda//sub j/ (equal to 0.849, 0.901, 0.908, 0.905, 0.897 for tone0, tone1, tone2, tone3, tone4 respectively) has been proposed. The calculation result showed that the relative length error of the proposed syllable duration method is 17.67%. An F0-contour generation approach to simulate the prosodic feature of a declarative sentence is also proposed in this paper. The preliminary hearing test showed that the intelligibility and the naturalness of synthetic speech were improved and achieved "good" level.