Speaker Recognition Based On Wavelet Packet And The Theory Of Chaos

Speaker recognition as a biological feature identification technology has important significance in the field of judicial authentication, information security, and human intelligence and so on. The new frequency band division method is studied in accordance with the principle of human ear perception. The feature extraction of speech signal is studied using the chaos technology. A chaotic characteristic model of speech is established to characterize the spatial motion state of the signal. The main contents are as follows:The wavelet packet decomposition based on Mel scale is proposed. The different requirements of speech signal for time and frequency domain was met by the flexibility of wavelet packet decomposition. An effective method of band division is provided by the Mel scale of human auditory perception integration to wavelet packet decomposition. Comparison with the traditional wavelet packet, the sub-band signal is divided by the wavelet packet based on Mel scale, which have more abundant information in time and frequency domain. The personality characteristics between different speech signals are highlighted.In order to solve the problem of chaotic decision of speech signal, the optimal delay time is determined by mutual information method. The embedding dimension is selected by the method of saturated correlation dimension. The phase space of speech time series is reconstructed. Properties of speech time series is determined using the maximum Lyapunov index method based on Wolf. The simulation results show that the speech signal is chaotic. In order to reflect the overall speech features, the internal structure characteristic of chaotic motion of the speech signal is characterized by fractal dimension. The box dimension feature is proposed to help other feature to improve the robustness of the speaker features.In order to solve the linear correlation and chaos of the speech signal, feature extraction scheme based on linear model and nonlinear model is proposed. In order to characterize the characteristics of low frequency energy, the auto regressive(AR) model is constructed based on wavelet packet in time and frequency domain. The acoustic model of speech chaotic time series is established using the Volterra adaptive prediction technique. The Volterra filter weight vector parameters are extracted as speaker features. The simulation results show that the prediction accuracy of this scheme is higher than the linear prediction model.The parameters of the AR model in time and frequency domain and the parameters of Volterra adaptive model are fused. Hidden Markov model(HMM) is used for speaker recognition. Simulation results show that the proposed feature extraction method can effectively improve the accuracy of speaker recognition, and achieve the desired results.