Variational auto-encoding generative model for speech to singing voice conversion

Voice conversion (VC) is a technique that involves replacing the identity of a person’s voice with another person’s voice without changing the speech content. This paper proposes speech to singing voice conversion using variational auto-encoding Wasserstein GAN (VAWGAN). The proposed method can separate the lyrical content and singer identity from the reference singing templates and later incorporate the target speaker identity in place of reference singer identity. In this way, this can generate a singing voice in the voice of the target speaker using reference singing templates. The performance can be assessed through subjective and objective evaluation experiments. In all the experiments the proposed VAWGAN method outperforms the baseline method. This can be performed well with limited non-parallel training data.