| Speaker recognition uses computer to identify the target speaker from speech sig-nals automatically.The dominant i-vector and x-vector systems have achieved great success.They map the given speech segment into a low-dimensional vector represen-tation which contains discriminative speaker information and the vector can be further processed through back-end modelling.Based on these systems,this paper studies the speaker recognition methods using low-dimensional vector representations,including the extraction methods,channel compensation algorithms and combination strategies for low-dimensional vector representations.The network structure is very important for the deep neural network(DNN)em-beddings to express the information about speakers.The original x-vector uses the time-delay neural network(TDNN)and the statistics pooling layer to extract and aggregate the frame-level information respectively,which is not enough to extract the speaker information accurately.The attentive statistics pooling(ASP)enhances the pooling ability of the neural network by weighting the high-level feature representations.In this paper,we further introduce the gating mechanism into the DNN embedding system.The gated convolution neural network(GCNN)is employed instead of TDNN to obtain more expressive frame-level feature representation.Besides,through the combination of the gating and attention mechanisms,we propose gated-attention statistics pooling(GASP)to further enhance the ability of capturing discriminative speaker information in the pooling layer.The x-vector network training is based on the supervised learning.During the training phrase,only the speaker label is used and the unsupervised information of the input features itself is neglected.In this paper,a multi-task learning strategy with high-order statistics is proposed for training the speaker embedding DNN.The network is trained with the primary task of classifying the target speakers,and the auxiliary task of reconstructing the first-and higher-order statistics of the original input utterance.The proposed training strategy aggregates both the supervised and unsupervised learning into one framework to make the speaker embeddings more discriminative and robust.The channel compensation algorithm can alleviate the channel information in the low-dimensional speaker vector.Although the linear discriminant analysis(LDA)is widely used for channel compensation,it is difficult to meet the requirements in complex environments due to some of the limitations identified for LDA,where the underlying distribution of classes is supposed to be Gaussian and unimodal.Supervised locality preserving projection(SLPP)preserves more important local information by weighing both the within-and between-speaker nearby data pairs based on the weight matrices.Compared with LDA and its variants(NDA and LWLDA),SLPP can achieve better performance.In this paper,combining the probabilistic linear discriminant analysis(PLDA)with SLPP,we propose P-SLPP to improve SLPP by introducing the relative PLDA score in the weight matrices.The P-SLPP algorithm can further enhance the ability of channel compensation for both i-vector based and x-vector based systems.A single low-dimensional vector contains limited speaker information.In this pa-per,we combine several low-dimensional vectors through score fusion,concatenating fusion and deep fusion,so as to explore the complementarity between diferent low-dimensional vectors.In the deep fusion,this paper proposes a fusion network based on AM-Softmax,which aggregates several low-dimensional vectors into one vector rep-resentation.The deep fusion achieves better performance than the score fusion and concatenating fusion.Moreover,it can be further improved through a multi-task learn-ing strategy.Finally,by combining the deep fusion and score fusion,the performance of speaker recognition system can be significantly improved. |
