Voice conversion based on Gaussian processes by using kernels modeling the spectral density with Gaussian mixture models

Voice conversion (VC) is a technique that aims to transform the individuality of a source speech so as to mimic that of a target speech while keeping the message unaltered. In our previous work, Gaussian process (GP) was introduced into the literature of VC for the first time, for the sake of overcoming the “over-fitting” problem inherent in the state-of-the-art VC methods, which gives very promising results. However, standard GP usually acts as somewhat a smoothing device more than a universal approximator. In this paper, we further attempt to improve the flexibility of GP-based VC by resorting to the expressive kernels that are derived to model the spectral density with Gaussian mixture model (GMM). Our new method benefits from the expressiveness of the new kernel while the inference of GP remains simple and analytic as usual. Experiments demonstrate both objectively and subjectively that the individualities of the converted speech are much more closer to those of the target while speech quality obtained is comparable to the standard GP-based method.