基于混合线性变换的语声转换算法

An Algorithm for Voice Conversion Based on Mixtures of Linear Transformation

针对在没有对称语音库的情况下,该文提出了一种基于混合线性变换的语声转换算法,在最大似然估计准则下,使用EM迭代算法计算变换函数的参量。为了减小线性加权对语音谱包络的平滑作用,使用线性调频Z变换来调节语音信号的LPC系数。客观评测和主观感受的实验结果都表明,基于混合线性变换的语声转换算法也可以取得与传统语声转换技术相当的转换效果,解除了传统语声转换技术需要对称语音库的要求。

This paper proposes an algorithm for voice conversion based on mixtures of linear transformation which avoids the need for parallel training corpus inherent in conventional approaches. In maximum likelihood framework the EM algorithm is used to compute the parameters of the transfer function. And the chirp Z-transform is utilized to enhance the smoothed spectral envelop due to the linear weighted averaging. The proposed voice conversion system is evaluated using both objective and subjective measures. The experiment results demonstrate that the proposed approach is capable of effectively transforming speaker identity and can achieve comparable results of the conventional methods where a parallel corpus is needed.