Reproduce “Training a Compact, Strong, and Realistic Text-to-Speech Model” 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
Runyuan Cai, Yu Lin, Yiming Wang, Xiaodong Zeng
Audio8-TTS adds a reference-anchored realism reward to zero-shot voice cloning. Its critic compares generated speech with a matched human recording. Listeners rated this stage higher in a blind study than earlier stages; CV3 error fell from 5.41% to 4.86% and speaker similarity rose from 0.639 to 0.661. The critic requires a matched reference and does not measure absolute naturalness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.