Modeling linguistic stress patterns in connected speech

The purpose of this paper is to present a model of linguistic stress patterns in connected speech. By categorizing the syllables roughly into three levels—unstressed, stressed, and prominently stressed—the magnitude of stress for the unstressed and stressed syllables can be accurately predicted, according to its relative position in the phrase group, knowing only the positions and the magnitude of a few prominently stressed syllables. The model shows that the perceived stress associated with each syllable is highly dependent on three factors: (1) the time lag between the present syllable and the last prominently stressed syllable (the recency effect), (2) the lead time between the same syllable and the forthcoming prominently stressed syllable (the anticipatory effect), and (3) the alternate stress rule. Data are collected from various speakers reading several chosen passages at different rates. The magnitudes of linguistic stress for these recordings are obtained both from perceptual judgments of trained listeners and from computer analysis. Results show that (1) stressed syllables occur at a regular rymthm, and (2) their magnitudes can be closely predicted by the model. Some implications of the present model and comparisons with other models are discussed.