We propose a speaker emotional state classification method that employs inference-based Bayesian networks to learn posterior density of emotional speech sequentially. We aim to alleviate difficulty in detecting medium-term states where the required monitoring time is longer compared to short-term emotional states that makes temporal content representation harder. Our inference algorithm takes advantage of the Sequential Monte Carlo (SMC) sampling and recursively approximates the Dirichlet Process Mixtures (DPM) model of the speaker state class density with unknown number of components. After learning the target posterior, classification of speaker states has been performed by a simple minimum distance classifier. Test results obtained on two different datasets demonstrate the proposed method highly reduces the training data length while providing comparable accuracy compared to the existing state-of-the-art techniques.