Reproduce “PhaseAug: A Differentiable Augmentation for Speech Synthesis to Simulate One-to-Many Mapping” 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
Junhyeok Lee, Seungu Han, Hyunjae Cho, Wonbin Jung
PhaseAug, a differentiable augmentation technique, improves the fidelity of GAN-based neural vocoders by rotating the phase of audio signals during training. This method simulates the inherent one-to-many mapping in speech synthesis, effectively reducing periodicity artifacts and boosting performance, especially in low-data scenarios.
Key insights
- Differentiable phase manipulation in the frequency domain can effectively simulate the one-to-many relationship of acoustic features to raw waveforms in speech synthesis.
- Applying such an augmentation reduces discriminator overfitting in GANs, providing more informative and less saturated gradients to the generator.
- The proposed method is architecture-agnostic, enhancing the performance of existing GAN-based vocoders without requiring structural changes, and demonstrates improved data efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.