To understand the process that the neural network has the classifying capability I calculate the condition for appropriate approximation of network function by reducing intrinsic entropy. Intrinsic entropy of a network is determined by the dimension of the network, i.e. the number of adaptible parameters. If the network can implement the desirable function, learning can be established easily by reducing the dimension of the network by simple architecture. This property is tested on the experiments of speech recognition using MLP and TDNN of various architectures and is verified to be correct. And the speech recognition experiments shows about 90% of recognition rates on the speaker dependent restricted sets of speech samples.