基于基元段特征和GMM的源-目标说话人F0～t转换

基音轨迹F_0～t转换是实现高质量源-目标说话人声音转换的重要组成部分。本文给出了一种与说话内容无关的F_0～t转换方法,为了在与文本无关的前提下,提取出尽可能反映说话人个性特征的基音起伏的较长时的信息,本文采用了从n个短时帧组成的基元段提取特征矢量,并以基元段矢量为单元进行转换,采用了基于高斯混合模型(GMM)的概率加权转换算法使每个特征矢量的转换规则是由多个类规则的线性加权组合得到的,从而提高了转换精度,同时还解决了一般分类器中处于类边界数据的分类错误。实验表明,基于基元段特征矢量和GMM的转换方法具有很好的效果。

F0～t Transformation From source Speaker To Target Speaker Based on Segment Feature Vector and GMM

The Transformation of F_0 contour from source speaker to target speaker is a key issue of the high quality voice trans- formation.In precondition of text-Independent,in order to extract the parameter which can reflect the speaker's relative long term pro- sodie characteristic feature,a kind of segment feature vector made up with n frames pitch is proposed as characteristic parameter to be transformed as a unit.This paper utilizes probability weighted transformation algorithm based on Gaussian Mixture Model(GMM), which linearly combines a few rules derived from each subclass transformation,thus the transformation accuracy is improved.Mean- while,for the features which nearly lie on the borders of several classes,the classification error may be eliminated.The proposed ap- proach is proved to be effective by conducted experiments.