Incremental speech understanding in a multimodal web-based spoken dialogue system

[摘要] In most spoken dialogue systems, the human speaker interacting with the system must wait until after finishing speaking to find out whether his or her speech has been accurately understood. The verbal and nonverbal indicators of understanding typical in human-to-human interaction are generally nonexistent in automated systems, resulting in an interaction that feels unnatural to the human user. However, as automatic speech recognition gets incorporated into web-based and portable interfaces, there are now graphical means in addition to the verbal means by which a spoken dialogue system can communicate to the user. In this thesis, we present a multimodal web-based spoken dialogue system that incorporates incremental understanding of human speech. Through incremental understanding, the system can display to the user its current understanding of specific concepts in real-time while the user is still in the process of uttering a sentence. In addition, the user can interact with the system through nonverbal input modalities such as typing and mouse clicking. We evaluate the results of a comparative user study in which one group uses a configuration that receives incremental concept understanding, while another group uses a configuration that lacks this feature. We found that the group receiving incremental updates had a greater task completion rate and overall user satisfaction.

[发布机构] Massachusetts Institute of Technology