A new hidden Markov model (HMM) is proposed for on-line character recognition using two typical features, pen-direction feature and pen-coordinate feature. These two features are quite different in their stationarity; pen-direction feature is stationary within every line segment of a stroke whereas pen-coordinate feature is not. In the proposed HMM, these contrasting features are used in a separative and selective way. Specifically speaking, pen-direction feature is outputted repeatedly at intra-state transition whereas pen-coordinate feature is outputted once at inter-state transition. The superiority of the proposed HMM over the conventional HMMs was shown through single-stroke and multi-stroke character recognition experiments.