In this paper, we present a new kind of neural network model based on subnets-SNN. The network is composed of a set of subnets and a decision layer. Each subnet is a Multilayer Perceptron that has two outputs units and used as a simple pattern classifier. The decision layer takes all the outputs of the subnets into account, and makes the final decision. We also give an training algorithm. Each pattern can be trained independently, which is possible to train a NN on a personal computer. In the paper, we compared SNN with the MLP, and show that SNN greatly decreases the compl exity of the networks. Evaluation experiments were conducted, using 10 Chinese vowel syllables. The results show that the SNN is effective and has more potential in the speech recognition.