The given English phonemes, words and sentences are sampled and preprocessed. For these real measured speech signal series, time delay and embedding dimension are determined by using mutual information method and Cao's method, respectively, so as to perform phase space reconstruction of the speech signal series. By using small data set method, the largest Lyapunov exponent of the speech signal series is calculated and the fact that its value is greater than zero presents chaotic characteristics of the speech signal series. This, in fact, performs the chaotic characteristic identification of the speech signal series. By introducing second-order Volterra series, in this paper we put forward a type of nonlinear prediction model with an explicit structure. To overcome some intrinsic shortcomings caused by improper parameter selection when using the least mean square (LMS) algorithm to update Volterra model efficiency, by using a variable convergence factor technology based on a posteriori error assumption on the basis of LMS algorithm, a novel Davidon-Fletcher-Powell-based second of Volterra filter (DFPSOVF) is constructed and is performed to predict speech signal series of the given English phonemes, words and sentences with chaotic characteristics. Simulation results under MATLAB 7.0 environment show that the proposed nonlinear model DFPSOVF can guarantee its stability and convergence and there are no divergence problems in using LMS algorithm; for single-frame and multi-frame of the measured speech signals, when root mean square error (RMSE) is used as an evaluation criterion the prediction accuracy of the proposed nonlinear prediction model DFPSOVF in this paper is better than that of the linear prediction (LP) that is traditionally employed. The primary results of single-frame and multi-frame predictions are given. So, the proposed DFPSOVF model can substitute linear prediction model on certain conditions. Meanwhile, it can better reflect trends and regularity of the speech signal series and fully meet requirements for speech signal prediction. The memory length of the proposed prediction model may be selected by the embedding dimension of the speech signal series. The proposed model can present a nonlinear analysis and more valuable model structure for speech signal series, and opens up a new way to speech signal reconstruction and compression coding so as to improve complexity and process effect of speech signal processing method.