Study Of Voiceprint Recognition Method Based On Improved Echo State Network

Voiceprint Recognition (VR) is also known as the Speaker Recognition (SR). Compared with other technology such as fingerprint recognition, palm print recognition, iris recognition and other biometric authentication technology, Voiceprint Recognition technology has attracted much attention because of its unique advantage on convenience, economy and accuracy. The voiceprint recognition models are commonly constructed as parametric models, non-parametric model, artificial neural network and support vector machine model. Among them, the artificial neural network has become the commonly used method in voiceprint recognition technology, because of its good classification, adaptive learning ability and strong fault tolerance.Echo State Network (ESN) is a new type of recurrent neural network, proposed by Jaeger in2001. It has attracted wide attention from academics and engineers owing to its unique reservoir structure and simple training algorithm, and has been widely applied in many fields, including the application in speech recognition.The traditional echo state network is connected by a random arrangement of large scale sparse (usually S) neurons to constitute a reservoir (hidden), and is used to complete the state mapping from low dimensional input space to high dimensional output space. The output layer is linear, which greatly simplifies the training process of weights. Studies have shown that, the choice of different types of neurons construct reservoir will lead to different impacts on network performance. In addition, the output layer using nonlinear readout method helps to make full use of high order statistics reserve output, and improves network performance.In this paper, we first construct a kind of improved LIF_ESN, by replacing the traditional S type neurons with Leaky Integrate and Fire Neuron. By the set of benchmark, the performance of LIF_ESN is superior to the classical ESN in the same reservoir size. Then this paper tries to use probabilistic neural network to read the high dimensional information of LIF_ESN’s reservoir in a nonlinear way, and provides another improved PLIF_ESN. Based on the baseline experiment, the performance of PLIF_ESN is slightly higher than LIF_ESN. Finally, based on the fact that the voiceprint recognition has a wide application prospect in the field of identity verification, this paper extracts four speaker voice signal characteristics with the MDLF_Mavg method, and uses ESN, the improved LIF_ESN and PLIF_ESN to identity speakers’voice signal characteristic value. Many groups of experiments show that, the two kinds of improved ESN have better performance than the traditional ESN when the reservoir size is limited. In addition, in our experiments, the ESN and the improved ESN obviously shows higher performance when compared with the traditional RBF network.