Regression approaches to voice quality controll based on one-to-many eigenvoice conversion

This paper proposes techniques for flexibly controlling voice quality of converted speech from a particular source speaker based on one-to-many eigenvoice conversion (EVC). EVC realizes a voice quality control based on the manipulation of a small number of parameters, i.e., weights for eigenvectors, of an eigenvoice Gaussian mixture model (EV-GMM), which is trained with multiple parallel data sets consisting of a single source speaker and many pre-stored target speakers. However, it is difficult to control intuitively the desired voice quality with those parametersbecause each eigenvector doesn’t usually represent a specific physical meaning. In order to cope with this problem, we propose regression approaches to the EVC-based voice quality controller. The tractable voice quality control of the converted speech is achieved with a low-dimensionalvoice quality control vector capturing specific voice characteristics. We conducted experimental verifications of each of the proposed approaches.