Reproduce “Score-Based Training for Energy-Based TTS Models” 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
Wanli Sun, Anton Ragni
The authors introduce delta loss for text-to-speech energy-based models, training the score to point from an acoustic model’s initial speech-feature estimate toward a reference. This targets useful refinement in one update without relying on contrastive negative examples. On 100 test clips, one-step Delta scored 3.95±0.05 MOS versus SSM’s 3.90±0.06; the close ratings do not establish superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.