Small Sample Voiceprint Recognition Based On Deep Learning

The globalization and application of contemporary technology,more and more occasions need to use identity recognition technology.Although the rapid development of the network brings convenience,it also brings hidden dangers of information leakage.Compared with other authentication technologies,the development of voiceprint recognition technology is not so mature.With the development of deep learning,voiceprint recognition technology relies on the training of deep neural networks to achieve the ideal recognition state.According to the definition of deep neural network,to make the network reach a good convergence state,it must have enough training samples,but it is very difficult to obtain enough voiceprint data.This paper is devoted to obtain a better recognition rate through small sample data.From the above point of view,this article conducts research according to the following:(1)First,this paper studies the concept and development history of voiceprint recognition technology.Secondly,it analyzes and learns the basic "deep learning" of voiceprint recognition technology in detail.Therefore,this paper proposes a basic training method based on small sample data set..(2)After the method is proposed,all the data sets are processed into spectrograms by the short-time Fourier transform technique for the data set,and the data set of the small sample data set is expanded to avoid the small number of samples,so that the network cannot reach a good convergence state;Secondly,select the convolutional neural network as the training network,train the large sample data set,and then transfer the convolutional layer to the small sample data set for training by transfer learning,and the fully connected layer is controlled by the restricted Boltzmann machine and Contrast with the divergence learning algorithm,and finally output all of them to the softmax classifier for calculation to obtain the mathematical model.(3)The actual operation of the above method,first verify the effectiveness of the source network,then verify the importance of transfer learning,and finally verify the performance of the above hybrid model through two different small sample data sets,thus determining the feasibility of the proposed method.