Unit feature based pruning of large-scale speech corpus using decision tree

Abstract

In this paper, we proposed and realized a corpus pruning method using decision tree. In the process of clustering, instead of conventional method, we measure the distance of pitch contours by feature vector composed by slope mean. The subjective and objective evaluation results showed that synthetic outputs based on corpus pruned through our method are close to outputs based on no-pruning corpus and are superior to conventional method with the same storage size.