In speech pattern recognition, there is a clear need to appropriately model the dynamics (variable durational nature) of pattern. This paper discusses a novel neural network solution to this requirement by proposing State-Transition Fuzzy Partition Model (STFPM). STFPM uses an HMM-like state transition structure, of which each state corresponds to one FPM network. The proposed network accordingly inherits all the advantages, such as a fast training and a robust decision, from the original FPM. Evaluations in speaker-dependent phoneme classification tasks clearly demonstrate the utility of this new network classifier.