사용자의 데스크 액티비티에 따른 스마트 스피커의 능동적

Background Users tend to more positively evaluate intelligent agents with higher personification properties. However, the conversations with smart speakers are currently initiated by the user's unilateral utterance of the call word, and the conversation does not take into account information or tastes of the user's situation. This differs from the general communication characteristics between people. In this work, we evaluate the user preference by different levels of active response of the smart speakers that perform proactive conversations by automatically recognizing the user's desk activity.Methods First, we defined a system concept based on deep learning and rule-based models with data acquired from microphones and sound sensors, human sensitivity sensors, and light sensors to automatically recognize the users' desk activity. Second, we divided the task situations that can be judged by the system into 19 different levels, and we derived specific scenarios by dividing the active level of the interaction into four stages: non-response, simple response, situation prediction and suggestion, and proactive response executing the suggestions. Third, we evaluated which level of active interaction is more preferred through user evaluation for each task situation.Results In most task situations, situation prediction and suggestion and proactive response interactions have been shown to be preferable to non-response, while simple response interactions have been evaluated negatively. In particular, the participants in the experiment were found to be concerned about context interruption, especially in situations where they were immersed in certain tasks or where there were several people together.Conclusions Smart speaker's proactive conversation depending on user's context will be very useful if the system’s higher recognition accuracy is supported, thereby providing a more extended user experience.