Improving GMM-based spectral conversion with optimal conversion function selection

We address the problem in the conventional Gaussian mixture model (GMM)-based spectral conversion from the viewpoint of optimal conversion function selection. The proposed method is motivated by that if the optimal conversion function based on minimum mel-cepstral distortion (MMCD) criterion can be selected during the conversion stage, the conversion performance in terms of mel-cepstral distortion (MCD) can be improved dramatically. To this end, our goal is to improve the accuracy rate of the optimal conversion function selection by the MMCD-based data clustering with Linear Discriminant Analysis (LDA). Experiment results confirmed that the proposed method can effectively improve the conventional method.