| The selection of acoustic features of speech signal and the way to train the deep neural network are crucial to improving the performance of speaker recognition system.This paper proposes a speaker recognition which based on the aggregate convolutional neural network(CNN)of feature guiding and multi-task learning.The overall work of this paper is as follows:1.The method which utilizes the time domain features guiding the CNN learn is studied,then the aggregate CNN of feature guiding is proposed for speaker recognition.Firstly,the filter bank(Fbank)feature of the speaker’s speech signal is extracted,and feed it into the CNN.Secondly,the statistic pooling layers are used to extract the mean vector and the standard deviation vector of the depth features of the output of each convolution layer,the mean vector and the standard deviation vector are spliced to obtain the statistic features of each convolution layer.Then,all the statistic features of the convolution layer are used as the input of the aggregate layer to get the aggregate feature.Finally,the guiding vector of the speech signal is concatenate with the aggregate feature to obtain the guiding feature,and the guiding feature as the input of the classier for speaker recognition.2.Combining the supervised learning and the unsupervised learning,a multi-task learning method is studied.Based on the CNN of multi-task learning is proposed for speaker recognition.Firstly,the Fbank feature is utilized as the input of the CNN.Secondly,the statistic pooling layers are utilized to extract the statistic features of the depth features of the output of each convolution layer,all the statistic features of the convolution layer as the input of the aggregate layer to obtain the aggregate feature.Then,the guiding feature is obtained by splicing the aggregate feature with the guiding vector of the speech signal.Finally,the guiding feature is used as the input of the fully connected(FC)layer and the Softmax layer.In this method,all of the networks before the Softmax layer are shared,and another FC layer is added to the first FC layer of the classifier to reconstruct the guiding vector of the speech signal.In the end,the CNN is trained by combining the cross-entropy loss and the reconstruction loss for speaker recognition. |
