188宝金博页面版

计算机技术与发展项目基金：国家自然科学基金（61401227）；江苏省博士后基金（1402067B）作者简介：崔立梅(1988—)，女，河北，硕士研究生，研究方向为语音转换；李燕萍（1983—），女，博士，副教授，研究生导师，通讯作者，主要研究方向为语音转换和说话人识别；吕中良（1990—），男，河南，硕士研究生，研究方向为语音转换。基于 ISODATA 聚类算法的语音转换研究崔立梅，李燕萍*，吕中良（南京邮电大学 通信与信息工程学院， 江苏 南京 210003）要 摘要 本文提出了一种基于迭代自组织聚类算法ISODATA的双线性频率弯折语音转换模型。根据语音特征参数分类不充分产生残差成分的问题，在基于高斯混合模型的聚类过程中引入了迭代自组织聚类算法。该算法将聚类得到的类内均值作为训练模型初始均值，改善了 EM算法初始值选取不当导致算法不能收敛的问题，从而对特征参数的拟合更加准确，结合后续的双线性频率弯折（BLFW）模型实现语音转换。实验测试结果表明：本文提出的算法具有较好的自适应聚类特性，能够使特征参数分类更合理，进而得到更准确的转换函数，使得转换的语音更接近目标语音。选择合适的初始值参数，本文提出的算法与高斯混合模型及双线性频率弯折模型相比较，平均 MCD 值相差很。骄 MOS 值提高。这说明合理精确的聚类有利于提高语音转换系统的性能。： 关键词： 迭代自组织聚类算法；双线性频率弯折语音转换模型；残差成分； 聚类特性Voice Conversion Based on Self Organizing ClusteringAndFrequency WarpingCui Limei, Li Yanping,Lv zhongliang( College of Communication & Information Engineering , Nanjing Universityof Posts and Telecommunications , Jiangsu , Nanjing 210003 , China)Abstract: This paper presents an Iterative Self-Organizing clustering Data Analysis Techniquesalgorithm ISODATA based on the bilinear frequency warping of the voice conversion model.According to the classification of speech feature parameters is not sufficient to generate residualcomponents, in the clustering process based on Gauss mixture model is introduced in the iterativeself-organizing clustering algorithm. The clustering algorithm is obtained within class averagevalue as the initial mean training model, and improves the initial value selection of EM algorithm,which leads to the problem that the algorithm can not converge, thus fitting the characteristicparameters are more accurate, with subsequent bilinear frequency warping (BLFW) model torealize voice conversion. The experimental results show that the proposed algorithm has betteradaptive clustering characteristics, which can make the characteristic parameters classificationmore reasonable, and get more accurate conversion function, making the speech more close to thetarget speech. Choosing appropriate initial value parameters, the algorithm proposed in this paperis compared with the Gauss mixture model and the bilinear frequency warpubg model, the averageMCD value is very small, and the average MOS value is high. This shows that reasonable andaccurate clustering is beneficial to improve the performance of speech conversion system.Keywords: the iterative self-organizing clustering algorithm ;the bilinear frequency warping voiceconversion model; residual components; clustering characteristicsclustering characteristics1 1 引言 语音包含很多信息，其中最主要的就网络出版时间：2017-04-28 17:02:45网络出版地址：http://kns.cnki.net/kcms/detail/61.1450.TP.20170428.1702.024.html