End-to-End Zero-Shot Voice Conversion Using a DDSP Vocoder

In this paper, we propose a zero-shot voice conversion algorithm using a neural vocoder based on differential digital signal processing. The vocoder does not require auto-regression, and its lightweight, differentiable nature allows the proposed system to be trained in an end-to-end fashion. This enables the use of more perceptually relevant objective functions for model training, and allows feature conversion and vocoder sub-networks to internally learn their own acoustic representation in a data-driven manner. We illustrate the effectiveness of the proposed algorithm by both qualitative and quantitative means, with comparisons to some of our previous works.