Reproduce “Phrase-Localized Language-Contrastive Guidance: Training-Free Localized Accent Control for Code-Switching Text-to-Speech” 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
Che Hyun Lee, Sangkwon Park, Donghun Kang, Dongwook Lee, Youngho Cho, Heeseung Kim, Sungroh Yoon
Phrase-Localized Language-Contrastive Guidance (LCG) offers a training-free solution for dynamic, localized accent control within code-switching Text-to-Speech systems. This approach leverages internal self-attention to achieve phrase-specific phonetic steering, resulting in an embedded phrase language accuracy of 0.518, a Mixed Error Rate reduction to 0.445, and 75.5% human preference for phrase nativeness while maintaining overall speech quality.
Key insights
- Dynamic phrase localization can be effectively achieved by adaptively probing internal self-attention mechanisms, with a strategy that prioritizes recall over precision for robust accent control.
- An independently tunable language-contrastive scaling factor (λ) allows for direct and stronger accent enforcement within specific phrase segments without destabilizing the global naturalness of the primary language carrier.
- Mask refinement strategies like dual-tag union and margin dilation enhance boundary recall, and the diffusion model's iterative denoising process absorbs transient over-dilation errors without propagating them to the primary language region.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.