Zero-shot singing voice conversion

In this paper, we propose the use of speaker embedding networks to perform zero-shot singing voice conversion, and suggest two architectures for its realization. The use of speaker embedding networks not only enables the capability to adapt to new voices on-the-fly, but also allows for model training on unlabeled data. This not only facilitates the collection of suitable singing voice data, but also allows networks to be pretrained on large speech corpora before being refined on singing voice datasets, improving network generalization. We illustrate the effectiveness of the proposed zero-shot singing voice conversion algorithms by both qualitative and quantitative means