Vector quantization-based parameter control for speech conversion with improved naturalness

Abstract

In this paper, a parameter control method is proposed to improve naturalness of converted speech. The proposed method first extracts feature parameters from both input and reference speeches per each frame, and then concatenates them into a feature parameter vector. Next, all the feature vectors are clustered into a certain number of representative vectors by using vector quantization. Finally, the vectors are used to control parameters for speech conversion. It is shown from the performance evaluation that the proposed method can convert speech with more naturalness than a linear parameter control method.