Metraipolis › archive › dc_fef20e1f720c
Pronunciation-Oriented Reinforcement Learning for Japanese Text-to-Speech with Kana-Domain ASR Rewards — arXiv:2610.07575
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.07575
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-07T21:22:48.351Z
- SHA-256 of the archived text
8ab84da0b36ba1434223692c719d5aa60ad60ae63a837b3276204ac810808ba5- Stable URL of this page
https://metraipolis.com/documento/dc_fef20e1f720c
How to cite
Pronunciation-Oriented Reinforcement Learning for Japanese Text-to-Speech with Kana-Domain ASR Rewards — arXiv:2610.07575 — arXiv (en), https://arxiv.org/abs/2610.07575 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fef20e1f720c (Metraipolis archive dc_fef20e1f720c, sha256 8ab84da0b36ba1434223692c719d5aa60ad60ae63a837b3276204ac810808ba5)
BibTeX
@misc{dcfef20e1f720c,
title = {Pronunciation-Oriented Reinforcement Learning for Japanese Text-to-Speech with Kana-Domain ASR Rewards — arXiv:2610.07575},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.07575},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fef20e1f720c; sha256 8ab84da0b36ba1434223692c719d5aa60ad60ae63a837b3276204ac810808ba5},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.07575. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Character error rate (CER) computed by automatic speech recognition (ASR) is widely used as an intelligibility reward for reinforcement learning (RL) post-training of text-to-speech (TTS) systems. For Japanese, however, orthographic CER introduces a representation mismatch for pronunciation-oriented optimization: distinct kanji readings may collapse to the same orthographic representation, while equivalent pronunciations may admit different orthographic forms. We instead compute CER in the kana domain using a kana-transcribing ASR model and reference readings (Kana-CER). Under matched group relative policy optimization (GRPO) conditions, Kana-CER reduces target-kanji reading error by approximately 26% relative to the orthographic CER reward, while maintaining comparable orthographic CER and similar speaker similarity and objective speech quality. It also reaches its best validation performance in substantially fewer optimization steps (4k vs. 18k). We further observe severe output elongation under unregularized Kana-CER optimization, which is substantially suppressed by KL regularization. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv eess.AS, nuovi lavori del giorno). Categorie arXiv: Sound (cs.SD) ; Audio and Speech Processing (eess.AS). Pagina del lavoro: https://arxiv.org/abs/2610.07575 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-07T20:54:04.266Z dalla pagina https://arxiv.org/list/eess.AS/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).