Different words tend to appear in different prosodic contexts. For each word, we would like to be able to characterize its typical prosodic contexts of appearance, and, for applications purposes, we would like to do this using probability distributions. This paper describes the phenomena and the need, and explains a way to use Gaussian Mixture Models for modeling prosody.
Index Terms: continuous probability distributions, prosodic features, language modeling, perplexity, lexico-prosodic linkages.