Pitch pattern clustering of user utterances in human-machine dialogue

The paper considers pitch pattern variations of user utterances in human-machine dialogue. For intelligent human-machine communication, it is essential that machines understand prosodic characteristics which imply a user's various attitude, emotion and intention beyond vocabulary. The authors' original focus is on particularly distinct pitch patterns and their roles in the actual dialogues. They used human-machine dialogues collected by a Wizard of OZ simulation. Many utterance segments belonged to clusters that were prosodically flat patterns. From the result, they considered that utterances which belonged to the other clusters and those which were far from the centroids included non-verbal information. In these utterances, there were talks to themselves and questions to the machine including emotional expressions of a puzzle or a surprise. These pitch patterns were not only rich in ups and downs, but also their slopes were upward, while the pitch patterns were generally even or a little downward. These results indicate that peculiar pitch period patterns show non-verbal expressions. In order to actually utilize such information on human-machine interactions, the representative pitch patterns should be investigated concerning their relationship to various types of communication.