An evaluation of many-to-one voice conversion algorithms with pre-stored speaker data sets

This paper describes an evaluation of many-to-one voice conversion (VC) algorithmsconverting an arbitraryspeaker’s voice into a particular target speaker’s voice. These algorithmseffectively generatea conversionmodel for a new source speaker using multiple parallel data sets of many pre-storedsource speakers and the single target speaker. We conducted experimental evaluations for demonstrating the conversion performance of each of the many-to-oneVC algorithms, including not only the conventional algorithmsbased on a speaker independentGMM and on eigenvoice conversion (EVC), but also new algorithms based on speaker selection and on EVC with speaker adaptive training (SAT). As a result, it is shown that an adaptation process of the conversionmodel improves significantlyconversion performance,and the algorithmbased on speaker selection works well even when using a very limited amount of adaptation data.