Voice conversion based on empirical conditional distribution in resource-limited scenarios

Abstract

In this paper, a computationally efficient voice conversion system has been designed in order to improve the performance in resource-limited scenarios. First, mixtures of Gaussians (MoGs) at fixed locations of Mel frequencies have been used to represent the spectrum of STRAIGHT compactly. Second, the key conditional distributions for prediction are approximated by building histograms of aligned features empirically. Experiments have confirmed that our proposed method can obtain fairly good results compared to the traditional method without huge computational costs.