基于ISODATA聚类算法的语音转换研究

摘要

A voice conversion model of bilinear frequency warping based on Iterative Self-Organizing clustering Data Analysis Techniques Algorithm (ISODATA) is put forward.According to the residual components generated by insufficient classification of speech feature parameters,in the clustering process based on Gaussian mixture model,the iterative self-organizing clustering algorithm is introduced.It takes average value within class obtained by clustering as the initial mean for training model,which improves the problem that the algorithm cannot converge due to inappropriated initial value selection of EM algorithm,thus making the characteristic parameters fitting more accurate,realization of voice conversion with subsequent bilinear frequency warping (BLFW) model.The experimental results show that the proposed algorithm has better adaptive clustering characteristics,which can make the characteristic parameters classification more reasonable,and get more accurate conversion function,making the speech more close to the target speech.Choosing appropriate initial value parameters,the algorithm proposed is compared with the Gauss mixture model and the bilinear frequency warping model.The average MCD value is very small,and the average MOS value is high.This shows that reasonable and accurate clustering is beneficial to improve the performance of speech conversion system.%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: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,对提出的算法与高斯混合模型及双线性频率弯折模型进行比较,平均MCD值相差很小,平均MOS值有所提高.这说明合理精确的聚类有利于提高语音转换系统的性能.