Automatic Motion Selection Method for Spoken Dialog Scenario Editor

Spoken dialog system is one of the most useful interface between human and machine. These systems need scenario data, which is represented by Finite State Transducer (FST). It is usually written by hand, but making it is a time-consuming job.In order to make a dialog scenario file easily, a new dialog scenario editor is developed. In dialog systems, it is important for natural communication that an agent speaks with natural gesture. However, it is difficult to make an appropriate motion for every utterance in a dialog scenario. In this paper, automatic motion assignment method is proposed.Most of motions have some meanings, and these meanings correspond to some keywords. Relationship between motions and keywords are investigated in advance, and an appropriate motion is selected based on corresponding keyword list. Moreover, the proposed motion selection method also considers sentence styles (interrogative, negative, and so on) of the main phrase in the sentence. Naturalness evaluation experiment was carried out for combinations between motions and utterances, and motion selection rules (keyword list and conditions) were made based on the results. Totally 83 keywords were listed for 31 motions.The performance of the motion selection method was evaluated. From this experiment, only 13.3% of input sentences were given an appropriate motion. However, many errors could be corrected by modifying the motion selection rules, and finally 80.0% of assigned motions were evaluated as “appropriate.”The dialog scenario editor with automatic motion assignment was developed. It deals with FST-based dialog system. In particular, it outputs a dialog scenario file for MMDAgent. This editor operated correctly, and a scenario file could be created easily.