This paper describes a speech pre-processing and feature extraction methods and described the principle of generalized regression neural network (GRNN). In order to use neural networks for speech recognition, this article uses the variable frame-shift average frame method to average the characteristic parameters of the collected voice frame, and the feasibility of the variable frame-shift average frame method in neural network input parameters normalization is verified by experiments. In this paper, according to this method, the speech recognition based on the generalized regression neural network (GRNN) successfully ported to an embedded system, and realized the pipe climbing robot’s real-time speech control.