基于GMM和概率修正码本的源-目标说话人声门波转换

Glottal Flow Transformation from Source Speaker to Target Speaker Based on GMM and Probability Correct Codebook

提出了一种用于源-目标说话人声门波导数参数转换的、基于勒让德正交分解的声门波导数波形参数提取方法。该方法将声门波导数波形在6维正交勒让德坐标系中的投影构成了描述其形状的特征矢量,并采用基于GMM的概率分类加权转换算法,使每个特征矢量的转换规则可由多个类所对应的规则的线性加权组合得到,可以使转换性能得到较大的提高。在此基础上,又给出了一种基于GMM的声门波导数波形的码本修正算法,以弥补声门波导数波形参数化而损失的含有说话人个性特征的高频送气分量和波纹分量。实验结果表明,本文方法转换性能明显好于基于矢量量化(VQ)的码本映射算法。

For high quality voice transformation, a novel parameter extraction scheme for glottal flow derivative is proposed based on Legendre orthogonal decomposition. The algorithm uses the six-dimensional Legendre orthogonal coefficients to form a vector for describing the shape of glottal flow derivative. Moreover, this paper utilizes probability weighted transformation algorithm based on Gaussian mixture model （GMM）, which linearly combines a few rules derived improve be lost codeboo from each subclass transformation, thus the transformation accuracy is significantly d. Furthermore, to model high frequency aspirated and ripple information, which may in the procedure of parameterization for glottal flow derivative, a probability correct k is used to compensate such information. Experimental results are proved to be effective.