粒子群优化径向基函数网络的语音转换  

Voice Conversion of Radial Basic Function Neural Network of Particle Swarm Optimization

作　　者：董添辉 张玲华 

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]江苏省通信与网络技术工程研究中心,江苏南京210003

出　　处：《计算机技术与发展》2017年第5期64-68,共5页Computer Technology and Development

基　　金：江苏省高校自然科学研究重大项目(13KJA510003);江苏高校优势学科建设工程(PAPD)

摘　　要：径向基函数神经网络具有结构简单和学习速度快等特点,因此常被用作语音转换的模型。隐层核函数的中心是影响径向基函数神经网络性能的重要参数,而传统的K-均值聚类算法受初值影响大,全局优化的效果不佳。所以,选择合适的优化算法来调整RBF网络核函数的中心参数,能改善整个网络的性能,从而提升语音转换的效果。而粒子群算法是一种基于迭代的优化算法,具有容易实现、算法参数少、收敛快和突出的全局寻优能力等特点。提出了一种改进的粒子群算法,优化了径向基函数的中心以提高网络性能,便于更准确地获得说话人与目标人之间谱包络的映射关系。实验结果表明,提出的方法能够有效提高神经网络的性能,使转换后的声音更接近于目标声音。Due to simple structure and fast learning, Radial Basis Function （RBF） neural network is used commonly in voice conversion system. The center of kernel function in hidden layer is the important parameter of influencing the RBF neural network, but traditional K -means clustering algorithm relies on the initial value, which is ineffective in global optimization. Therefore,it is significance to select a suitable algorithm to modulate the center of function and enhance the effect of voice conversion. Particle swarm algorithm is an optimized one based on iteration, with the characteristics of easy implementation, much less parameters, fast convergence and better global optimiza- tion and so on. An improved particle swarm optimization is proposed to optimize the RBF＇ s centers for improvement of the performance of RBF network, thus enhancing the transformation of speech parameters. The results acquired by modeling and simulation show that the proposed method has effectively improved the performance of neural network and the effect of converted voices is much closer to the goal.

关 键 词：语音转换 径向基函数中心 改进的粒子群算法 径向基函数神经网络 

分 类 号：TN912.3[电子电信—通信与信息系统]