A Prediction Model for End-of-Utterance Based on Prosodic Features and Phrase-Dependency in Spontaneous Japanese

This study aims to reveal a clue for predicting end-of-utterance in spontaneous Japanese speech. In casual everyday conversation, participants must predict the ends of utterances of a speaker to perform smooth turn-taking with small gaps or overlaps. Syntactic and prosodic factors are considered to project the end of utterance of speech, and participants utilize these factors to predict the end-of-utterance. In this paper, we focused on the dependency structure among bunsetsu-phrases as a syntactic feature and F0, intensity, and mora duration for bunsetsu-phrases as prosodic features. We investigated the relationship between the position of a bunsetsu-phrase in an utterance and these features. The results showed that a single feature cannot be an authoritative clue that determines the position of bunsetsu-phrases. Next, we constructed a Bayesian hierarchical model to estimate the bunsetsu-phrase position from the syntactic and prosodic features. The results of the model indicated that prosodic features vary in usefulness according to speakers. This suggests that the different combinations of syntactic and prosodic features for each speaker are relevant to predict the ends of utterances.