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

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.

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

In order to impro ve the ef fect of vo ice conv ersion and reduce the data quant ity, an algo rithm of voice co nv ersion metho d w as pr opo sed based on Fo ur ier-Bessel ex pansion and K -L ( Karhunen-Loeve ) t ransfo rm, w hose transform model is GMM. This mo del of alg orithm appro aches the rule of vo ice t ransmission and can describe voice sig nals in vivid detail. In the alg orithm, a f rame-divided method w as put fo rw ard based on maximum phase of fundamental frequency and a parameter -compressed method based on K-L t ransform to aim at a bet ter ef fect . The co rresponding emulat ional experiment s w ere carried out , and the resul t of the ABX test to the t ransformed signal show s that the algo rithm performs well