We present the empirical evaluation of a probabilistic model of student affect based on Dynamic Bayesian Networks and designed to detect multiple emotions. Most existing affective user models focus on recognizing a specific emotion or lower level measures of emotional arousal, and none of these models have been evaluated with real users. We discuss our study in terms of the accuracy of various model components that contribute to the assessment of student emotions. The results provide encouraging evidence on the effectiveness of our approach, as well as invaluable insights on how to improve the model’s performance.