基于2DPCA特征降维的CNN说话人识别

Speaker Recognition Based on 2DPCA Feature Dimension Reduction and CNN

针对使用话语级特征参数矩阵作为卷积神经网络输入而导致收敛速度慢及识别率低的问题,提出一种基于二维主成分分析(2DPCA)特征降维的卷积神经网络(CNN)说话人识别方法。首先将每段语音分帧成多个帧级语音并提取同等大小的帧级特征组成特征矩阵,然后利用2DPCA对特征矩阵进行降维处理,再将得到的主成分特征向量组合成新的特征矩阵作为CNN的输入,最后通过CNN的自适应特征学习创建说话人模型。基于Alexnet的CNN模型实验结果表明,采用该说话人识别方法使运行时间减少了57%,同时识别率也有所提高。

Aiming at the problem of slow convergence and low recognition rate caused by using utterance-level feature parameter ma⁃trix as input of convolutional neural network.A speaker recognition based on 2DPCA(two dimensional principal component analysis)feature dimension reduction and CNN(convolutional neural network)was proposed in this paper.Firstly,each speech was divided into several frame-level speech,and the frame-level features of the same size were extracted to form the feature matrix.Then,2DPCA was used to reduce the dimension of the feature matrix,and the principal component feature vectors were combined into a new feature ma⁃trix as the input of CNN.Finally,the speaker model was created through adaptive feature learning of CNN.The experimental results with CNN model based on Alexnet show that the running time is reduced by 57%and the recognition rate is improved after using the speaker recognition method.