摘要
EM算法是高斯混合模型参数估值的常用方法,该算法有局部收敛的特性,易造成模型的参数估计对于初值较为敏感,往往得到一个局部的最优值。为了对EM算法进行优化,文中将具有全局寻优和并行搜索特性的遗传算法与EM算法相结合,对其加以改进,并用到语音转换过程之中,最后通过仿真实验分析了算法的性能,结果表明使用优化算法得出的高斯混合模型所转换出来的语音,相对于传统EM估计算法得出的高斯混合模型所转换出来的语音,具有较小的失真测度值,证明使用该优化算法能够改善转换后的语音质量。
EM algorithm is a common method to estimate the parameters of GMM. For its local convergence property,the EM algorithm is sensitive to the initial values and consequently lead to a subprime value. In order to optimize EM algorithm,combine the genetic algo- rithra with EM algorithm to improve it. Apply genetic algorithm parallel search and global optimization characteristics to voice conversion process. Compared with the traditional EM algorithm,the simulation results show that the improved algorithm has a small distortion measure values. So, the proposod method can improve the converted voice quality.
出处
《计算机技术与发展》
2011年第11期145-148,共4页
Computer Technology and Development
基金
南京工程学院高教研究重大课题(GY200802)
