Ruby-OPD enhances pronunciation controllability for flow-matching TTS
Original title: Ruby-OPD: Enhancing Pronunciation Controllability and Accuracy for Flow-Matching Text-to-Speech via On-Policy Distillation
Ruby-OPD framework improves pronunciation accuracy on raw orthographic text for Japanese flow-matching TTS.
Achieving high pronunciation accuracy and fine-grained controllability remains a central challenge in Japanese text-to-speech. Traditional cascaded systems discard lexical semantics and prosodic cues. Recent zero-shot TTS models suffer from polyphone errors and lack explicit control. We propose a two-stage framework for flow-matching TTS using Ruby text representations with orthography perturbation and On-Policy Distillation on large-scale text-only corpora. On the Joyo-Kanji-Yomi benchmark our model achieves 97.13% Accuracy, 2.02% Target PER, 0.70% Sentence PER and 4.21% Sentence CER under raw text while retaining strong controllability.
- Two-stage framework preserves surface forms and target pronunciations
- On-Policy Distillation enhances accuracy on raw orthographic text
- Benchmark results: 97.13% Accuracy, 2.02% Target PER
- Strong controllability retained alongside accuracy gains
From the original article
1 min read at arxiv.orgAchieving both high pronunciation accuracy and fine-grained controllability remains a central challenge in Japanese text-to-speech (TTS), as Japanese is an unsegmented language mixing logographic Kanji and phonographic Kana with extensive polyphony.
By Rui Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.