Research On Speaker Recognition Method Based On Fuzzy Neural Network

The speaker recognition technology authenticates the identity of the speaker by the extraction of the phonetic parameters which can characterize the biological features of the speaker.In recent years,due to the in-depth study of artificial intelligence technology by researchers,speaker recognition technology has also developed rapidly.Speech researchers have gradually shifted the focus of work from the construction of the speaker recognition system based on the Gaussian mixed model to the establishment of the speaker recognition system based on neural networks.Fuzzy neural networks possess a relatively strong capability in the model classification of phonetic signal of the speaker.Convolutional neural networks are capable of deep mining of the data features and moderate dimensionality reduction of the data.This thesis studies the speaker recognition effect by proposing an improved speaker recognition method.The main content of the thesis is as follows.(1)Aiming at how to improve the ablity of deep mining of data features by fuzzy neural network,this thesis effectively combines the convolutional layer and pooling layer of convolutional neural network with fuzzy neural network,and proposes a speaker recognition method based on improved fuzzy neural network.The improved fuzzy neural network utilized the depth of the convolutional layer to extract the data features and used the pooling layer to reduce the dimension of the feature data.After fuzzification of the output data of the pooling layer,the fuzzy inference system would calculate the fitness of the current rule according to the membership value of the input signal.After data defuzzification,the output layer would generate the speaker recognition result.(2)In view of the problem that the current Dropout algorithm may cause loss of key information due to random discarding of neurons,this paper improves the Dropout algorithm.In the current Dropout algorithm,a mechanism for judging and classifying the threshold values of neuron output values is introduced,and the possibility of discarding neurons with higher activation is reduced by the custom function.The experimental results show that the improved Dropout algorithm can effectively improve the recognition performance and generalization ability of the speaker recognitionsystem.