Vocal tract spectrum transformation based on clustering in voice conversion system

摘要

By the conventional vocal tract spectrum transformation based on Gaussian Mixture Model (GMM), the transformation rule is not very accurate because of the large amount of voice data which is time-varying and non-stationary. This paper mainly studies a method of spectrum transformation based on clustering algorithm. First of all, the training data are classified into several clusters and each cluster is trained relatively to get a more accurate transformation rule. And in the stage of transformation, the source parameters of each frame are classified into one cluster, and then are converted by the transformation rule of that cluster. In this paper, K-means algorithm is used as the clustering method to classified data. Experiment results show that proposed method based on clustering is better than the transformation by conventional GMM, especially the one by K-Means algorithm with 20 centers is the best one.