This work proposes a cloud system, structured as a set of REST API endpoints, for online human emotion recognition in spontaneous human-robot verbal interaction. Based on the appraisal theory of emotion, the system acquires data about the person's expected appraisal of a given situation, depending on their needs and goals, and combines it with sensory data, such as facial expressions, angles of the head, and gaze of the person, and distance between the person and the robot. The whole set of data is used to infer the emotional state of the person during the interaction through a Random Forest classifier, trained for binary classification (i.e., positive vs. negative emotions). Results confirmed that using both sources of data led to a performance improvement both in the K-fold and in the Leave One Person Out scenarios.