With rapid development of voice control technology, making speech recognition more precisely in various IoT domains have been an intractable problem to be solved. Since there are various conversation scenes, understanding the context of a dialog scene is a key issue of voice control systems. However, the reality is available training data for dialog system are always insufficient. In this paper, we mainly solve the problem of data lacking in dialog systems by data augmentation technique. A Generative Adversarial Network(GAN)-based model is proposed and the data are augmented effectively. It can generate from text to text, enhance the original data with text retelling, and improve the robustness of parameter estimation of unknown data by using the sample data generated by GAN model. A new N-gram language model is used to evaluate multiple recognition candidates of speech recognition, and the candidate sentences with the highest evaluation scores are selected as the final result of speech recognition. Our data enhancement algorithm based on the Generative Model is verified by the experiments. In the result of model comparison test, the error rates of data set THCHS30 and AISHELL are 3.3% and 5.1% which are lower than that of the baseline system.