Research On Speaker Recognition Base On The Ensemble Deep Learning Model

With the continuous progress of human society and the rapid development of computer technology,traditional identity authentication is gradually difficult to meet people’s growing needs for security,convenience and other aspects,and biometric authentication has been widely popularized and accepted in the society.Speaker recognition,also known as voice print recognition,is a kind of biometric feature authentication technology to identify the speaker’s identity according to the personality information contained in the voice signal.Compared with other biometric authentication methods such as fingerprint recognition and iris recognition,speaker recognition has the advantages of lower voice collection cost,weaker privacy and more convenient recognition methods.Therefore,speaker recognition technology is easier to be accepted by people and plays an important role in every field of human society.In this paper,the speaker recognition method based on DenseNet network structure is mainly studied for the text independent speaker recognition task which is widely applied.The network structure is improved on the basis of DenseNet,and coordinate attention mechanism and integrated learning theory are introduced into the speaker recognition field.A speaker recognition method based on D-DenseNet,D-Densenet-CA and integrated model are proposed.The main work of this paper is as follows:1.Improve the DenseNet network structure.DenseNet establishes dense connections between the current layer and all previous layers by means of feedforward neural network to enhance feature transmission,and realizes feature reuse by using the connections of features on channels,thus achieving better performance with fewer parameters.However,the DenseNet structure only establishes dense connections between dense layers inside each dense block,and does not take into account the feature transfer between dense blocks.Based on DenseNet network,this paper proposes a speaker recognition method based on D-DenseNet network.By establishing dense connections between dense blocks in DenseNet structure,the feature transmission in the network is further enhanced,and the overall feature extraction capability of the network is improved.2.Improve the correlation between channels for features that cannot be obtained by convolution operation in the network.The feature map extracted from the network layer contains different feature information in different channels,and the contribution of these feature information to the overall recognition performance of the network model is different.However,conventional convolution operation cannot obtain the dependency relationship between the features in the channels.In this paper,CA attention mechanism and D-DenseNet network are introduced to study a speaker recognition method based on D-Densenet-CA network.CA attention mechanism is used to capture the dependency between channels of feature maps and orientation related location information,so as to enhance the feature representation ability of the network model and improve the recognition performance of the speaker recognition system.At the same time,considering that a single model may be difficult to perform comprehensive performance in complex speaker recognition tasks,ensemble learning theory is introduced to construct three individual network models with different structures,namely DenseNet,D-DenseNet and D-Densenet-CA,and the three models are combined into an integrated model by weighted voting method and strategies.The integrated model can adjust the prediction results of the individual model,further reducing the probability of model prediction error,so as to obtain higher recognition accuracy.