基于深度学习及核典型相关分析的多特征融合说话人识别  

Speaker Recognition Based on Multi-Features Fused by Deep Learning and Kernel Canonical Correlation Analysis

作　　者：卜禹 陆璐璐 BU Yu;LU lulu(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212000)

机构地区：[1]江苏科技大学计算机学院

出　　处：《计算机与数字工程》2019年第9期2185-2189,2205,共6页Computer & Digital Engineering

摘　　要：论文提出一种基于深度学习以及核典型相关分析(kernel canonical correlation analysis,CCA)的多特征融合说话人识别方法。针对说话人的音频和视频信息,利用深度信念网络和卷积神经网络这两种不同深度的神经网络对音频信息和视频信息分别并行处理,得到两种不同模态的生物特征向量。采用核典型相关分析方法对这两种非线性相关的特征向量进行特征级融合,使用它们的相关性判别函数抽取多个相关性顺次下降但又互不相关的典型变量对按照给定的特征级融合策略构成最后的判别特征,同时去除了冗余信息。最后生成的基于核典型关联分析的融合特征输入最近邻分类器,输出说话人识别结果。使用BANCA数据库对该方法进行实验,结果表明:该方法能显著提高说话人识别的准确率。The method of speaker recognition based on multi-features fused by deep learning and kernel canonical correlation analysis is proposed in this paper. To acquire two biological features of different modes,deep belief network and convolutional neural network are used to respectively process the audio information and the video information of speaker in parallel. This two non-linear correlated features are fused in feature level by kernel canonical correlation analysis method,the correlation discriminant function is used to extract several pairs of canonical correlation variables whose correlation decrease sequentially but not correlated to each other to constitute the final discriminant feature as the given fusion strategy,which removes redundant information at the same time. At last,the fused feature generated by the kernel correlation analysis method is input into the K-nearest neighbor classifier,and the result of speaker recognition is output from the classifier. BANCA database is used to test the method proposed in this paper,the result shows that this method can improve the accuracy of speaker recognition significantly.

关 键 词：深度信念网络 卷积神经网络 核典型关联分析 最近邻分类器 说话人识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]