FastVoiceGrad: One-step Diffusion-Based Voice Conversion with Adversarial Conditional Diffusion Distillation

TL;DR

FastVoiceGrad introduces a one-step diffusion-based voice conversion method that significantly speeds up inference while maintaining high speech quality and speaker similarity, outperforming previous multi-step models.

Contribution

The paper presents FastVoiceGrad, a novel one-step diffusion-based voice conversion model using adversarial conditional diffusion distillation, reducing inference time drastically.

Findings

Achieves superior or comparable VC performance to multi-step models.

Significantly reduces inference time to a single diffusion step.

Demonstrates effective one-shot any-to-any voice conversion.

Abstract

Diffusion-based voice conversion (VC) techniques such as VoiceGrad have attracted interest because of their high VC performance in terms of speech quality and speaker similarity. However, a notable limitation is the slow inference caused by the multi-step reverse diffusion. Therefore, we propose FastVoiceGrad, a novel one-step diffusion-based VC that reduces the number of iterations from dozens to one while inheriting the high VC performance of the multi-step diffusion-based VC. We obtain the model using adversarial conditional diffusion distillation (ACDD), leveraging the ability of generative adversarial networks and diffusion models while reconsidering the initial states in sampling. Evaluations of one-shot any-to-any VC demonstrate that FastVoiceGrad achieves VC performance superior to or comparable to that of previous multi-step diffusion-based VC while enhancing the inference speed. Audio samples are available at https://www.kecl.ntt.co.jp/people/kaneko.takuhiro/projects/fastvoicegrad/.