NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:3681
Title:Blow: a single-scale hyperconditioned flow for non-parallel raw-audio voice conversion

the paper proposes a normalizing-flow based generative model called Blow for non-parallel raw-audio voice conversion that makes use of speaker identity labels. The model makes use of a single-scale structure, conditioning module based on hypernetworks, shared speaker embeddings. Reviewers have recognized the novelty of the paper that consider normalizing flows for voice conversion. They also liked the extensive experiments as well as the ablation studies, and the provided samples of the voice conversions.