Generation of pauses within the z-score model

We have previously proposed BB94] a model for the generation of segmental durations which proceeds in two steps : (a) prediction of the timing of a salient acoustic event per syllable according to phonotactic and syntactic information, and (b) application of a repartition model which determines the duration of each individual segment between these events. This paper focusses on the repartition model and describes how the initial model has been enriched to account for the emergence of pauses as speech rate is decreased. The last section of this paper describes a perceptual evaluation of the whole model. This evaluation shows that, for the same distribution of prediction errors, a precise timing of these events is perceptually more relevant than a segment-based method aiming at predicting precisely each individual segmental duration.