Reproduce “Pronunciation-Oriented Reinforcement Learning for Japanese Text-to-Speech with Kana-Domain ASR Rewards” 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
Shiao Zhu, Lianbo Liu, Kai Washizaki, Koki Nikaido, Yui Sudo
The authors test whether kana-domain ASR rewards improve Japanese TTS pronunciation accuracy by comparing generated speech with intended kana readings during GRPO training. The selected Kana-CER-reward model reaches 7.17±0.07% target-reading error, versus 9.65±0.12% with orthographic CER reward, a 25.7% relative reduction; orthographic CER remains comparable at 4.15±0.04% versus 4.10±0.16%. The result applies to one 0.5B base model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.