Reproduce “MeanVoiceFlow: One-step Nonparallel Voice Conversion with Mean Flows” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Takuhiro Kaneko, Hirokazu Kameoka, Kou Tanaka, Yuto Kondo
NTT, Inc. introduces MeanVoiceFlow, a one-step nonparallel voice conversion model that achieves high-quality, speaker-similar speech without requiring multi-step inference, pretraining, or knowledge distillation. The model demonstrates performance comparable to or surpassing multi-step and distillation-based methods on VCTK and LibriTTS datasets, with a specific pMOS_v improvement to 4.10.
Key insights
- Mean flows provide a mechanism for accurate direct computation of path integrals, allowing for effective one-step generative sampling without compromising quality.
- A carefully designed structural margin reconstruction loss applied as a zero-input constraint offers a robust regularization strategy, avoiding over-smoothing while guiding model learning.
- Aligning the training and inference input distributions through conditional diffused-input training is crucial for achieving stable, robust, and high-quality results in nonparallel voice conversion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.