An empirically based probabilistic model of intonation in American English

In this dissertation the phonology of American English intonation is analyzed using statistical tools. I propose a probabilistic model of intonation that is based on a large corpus and addresses in detail how tones combine. The model incorporates the following findings. Pitch accents do not occur in free variation, but have certain probabilities of occurring subsequent to one another. The model distinguishes between pitch accents occurring in non-final versus final intermediate phrases. The last pitch accent in the final intermediate phrase is found to be a strong predictor of boundary tone; this is modeled as a second order Markov process. This suggests that the tones in a tune are interrelated in a way that a model that assigns separate meaning to each tone cannot capture. The data are modeled by the assumption of a constant probability that a pitch accent will be followed by an additional pitch accent rather than the phrasal tone; such a model provides a good fit. Similar arguments hold for the number of intermediate phrases in an intonational phrase. L+H* is included as a phonological primitive; speakers tend to use L+H* in phrases that end with a fall-rise (L+H*L-H%) but use H* in phrases that end with a fall (H*L-L%).; I remove downstep from the phonological inventory based on careful statistical study. In my data I find no distinction between tones labeled downstepped and tones labeled nondownstepped. If tones considered downstepped are indeed in a compressed pitch range relative to the preceding tone, then they should have a target drop in frequency that is different from the target for nondownstepped tones. Instead I find that the frequency drops obey a normal distribution that shows no evidence of bimodality. The values of the frequency drops for downstepped and nondownstepped tones form the right- and left-hand tails of this distribution. I also present data indicating that steep rises in frequency may occur in peak accents immediately preceded by accents labeled as downstepped. Such rises are not explained by current intonational theory, which holds that a compressed pitch range is maintained until an intermediate phrase boundary.