基于仿射传播聚类的说话人识别算法  

Research on speaker recognition algorithm based on affinity propagation clustering

作　　者：张辰 张华 高宁化 陈豪 ZHANG Chen;ZHANG Hua;GAO Ninghua;CHEN Hao(Key Laboratory of Robot Technology Used for Special Environment of Sichuan Province,School of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,China)

机构地区：[1]西南科技大学信息工程学院特殊环境机器人技术四川省重点实验室

出　　处：《传感器与微系统》2020年第2期120-123,共4页Transducer and Microsystem Technologies

基　　金：四川省科技创新苗子工程资助项目(2018047,2017021)

摘　　要：为了提高说话人识别系统的离线训练效率,提出了一种基于仿射传播(AP)聚类的说话人识别方法,通过仿射传播聚类对说话人声纹特征样本进行样本筛选,采用神经网络算法训练说话人识别分类器,完成说话人识别。在自采集数据集上进行了说话认识别实验,证明采用仿射传播聚类算法对样本集进行大幅压缩过后,样本压缩率约为82%,网络训练时间下降率约为86.99%,而识别准确率与压缩前的识别准确率基本一致。实验证明了仿射传播聚类与神经网络结合的说话人识别算法可以在保证识别率的同时大大提高神经网络的训练速度,节约时间成本。In order to improve the offline training efficiency of speaker recognition system,a speaker recognition method based on affinity propagation clustering is proposed.The method uses affinity propagation(AP)clustering for screening sample of voiceprint feature of speakers.Then,the speaker recognition classifier is trained by neural network algorithm.Finally,speaker recognition is completed.In order to verify the effectiveness of the method,speaker recognition experiments are conducted on self-collected data sets.It is proved that the sample compression rate is about 82%and the training time of the network is about 86.99%after the sample set is compressed by affinity propagation clustering algorithm.The recognition accuracy is basically the same as that before compression.Experiments show that the speaker recognition algorithm based on affinity propagation clustering and neural network can greatly improve the training speed of neural network and save time and cost while guaranteeing the recognition rate.

关 键 词：说话人识别 仿射传播聚类 神经网络 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置] TP391.98[自动化与计算机技术—控制科学与工程]