Making smart speakers proactive might benefit users by delivering unconscious but useful information. However, depending on the timing and manner of the delivery, proactive deliveries pose a risk in disturbing user’s ongoing activity. Thus, we hypothesized that the times of activity-transition are appropriate for information delivery. In this study, we prototyped a system that estimates activity-transitions based the user’s body motion using a depth camera, and actively provide Internet news through a voice of virtual character. We then conducted a set of experiments in everyday-living scenario at the homes of five university students living alone. The result demonstrated that the times after transfer are generally appropriate. In contrast, the change of the user’s face direction, i.e. supposed gaze-target transition, occurred more frequently but included more inappropriate cases. Elimination of the error-detections such as body posture change while looking at smartphone is needed to detect activity-transitions more appropriately.