Modeling Pronunciation Variation in Conversational Speech Using Prosody

A significant source of variation in spontaneous speech is due to intra-speaker pronunciation changes. Previous work in automatic speech recognition has identified several factors that affect pronunciation variability such as phonetic context and speaking rate, as well as syntactic structure. This work examines prosody as a cue to pronunciation variability, as represented by attributes derived from F0, energy and duration values. Analyses of hand-labeled data are used to determine useful instances of prosodic variables for characterizing pronunciation changes, which in turn are used in a decision-tree-based dynamic pronunciation model. Experiments predicting phone changes show an improvement over chance when prosodic attributes are used. Including prosodic variables in a model using phonetic context and word-based information shows a 14% reduction in entropy and a slight improvement in phone error rate over the baseline model.