基于Fourier-Bessel展开和Karhunen-Loeve变换的语音变换  

Voice transform method based on Fourier-Bessel expansion and Karhunen-Loeve transform

作　　者：岳振军[1] 宋巍[2] 王浩[1] 张雄伟[2] 

机构地区：[1]解放军理工大学理学院,江苏南京211101 [2]解放军理工大学通信工程学院,江苏南京210007

出　　处：《解放军理工大学学报（自然科学版）》2008年第1期15-19,共5页Journal of PLA University of Science and Technology(Natural Science Edition)

摘　　要：为了有效地进行语音变换,改善变换后语音的自然度和目标人倾向度。依据语音信号传播机理和(Fourier-Bessel)展开式系数对语音信号的表现能力,提出了利用F-B展开系数作为变换参数。在该算法中,根据F-B展开系数无语音相位信息的特点,提出基于最大基频相位的语音分帧算法;针对F-B展开式数据量过大的问题,提出了基于Karhunen-Loeve变换的参数压缩算法,转换模型使用GMM(Gaussian mixture model)模型。对算法进行了仿真实验。对变换后语音所进行的ABX测试表明,算法能够较好地完成语音变换,变换后语音的目标人趋向度比较高。In order to improve the effect of voice conversion and reduce the data quantity,an algorithm of voice conversion method was proposed based on Fourier-Bessel expansion and K-L（Karhunen-Loeve） transform,whose transform model is GMM.This model of algorithm approaches the rule of voice transmission and can describe voice signals in vivid detail.In the algorithm, a frame-divided method was put forward based on maximum phase of fundamental frequency and a parameter-compressed method based on K-L transform to aim at a better effect.The corresponding emulational experiments were carried out,and the result of the ABX test to the transformed signal shows that the algorithm performs well.

关 键 词：语音变换 傅里叶—贝塞尔展开 KARHUNEN-LOEVE变换 高斯混合模型 

分 类 号：TN973.3[电子电信—信号与信息处理]