This paper studied Mandarin syllables recognition dynamic neural network called state detector neural network (SDNN), which had minimum number of state detectors. We described the speech signal with a sequence of states and each state statistically characterized a period of speech vector. In each state detector which constituted the dynamic neural network, input vector nonlinearly approximated to a desired vector that was based on a forcing learning algorithm that has the merit of less training epoch and higher convergence accuracy. We compare this model to three other neural network (NN) models including the model of NN/HMM-based warping on the same database. Experiments show that our method has higher recognition accuracy than the others.