Ruby-OPD
A two-stage flow-matching TTS model for Japanese text-to-speech. It achieves pronunciation controllability through ruby text representation and orthographic perturbation, and combines On-Policy Distillation to improve the accuracy of polyphonic kanji readings without paired speech data.
Timeline (last 90 days)
Ruby-OPD proposes a two-stage framework aimed at simultaneously resolving the tension between pronunciation controllability and naturalness in Japanese TTS on flow-matching TTS architectures.
Ruby text is a typographic convention in Japanese in which reading annotations are placed above kanji; Ruby-OPD uses it as a representational-layer device, using orthographic perturbation so that the model is exposed to both surface written forms and target pronunciations during training.
Ruby-OPD introduces On-Policy Distillation (OPD) to continue optimizing pronunciation accuracy on large-scale text-only corpora without paired speech.
On the Joyo-Kanji-Yomi benchmark, Ruby-OPD achieves an accuracy of 97.13%, a Target Phoneme Error Rate (Target PER) of 2.02%, a Sentence-level Phoneme Error Rate (Sentence PER) of 0.70%, and a Sentence-level Character Error Rate (Sentence CER) of 4.21% under raw text input.
The acquisition cost of Ruby annotations, the design details of the perturbation strategy, and the stability of OPD on larger-scale corpora remain engineering issues that need to be observed in follow-up work.
All Facts (5)
Ruby-OPD proposes a two-stage framework aimed at simultaneously resolving the tension between pronunciation controllability and naturalness in Japanese TTS on flow-matching TTS architectures.
50%UnverifiedRuby text is a typographic convention in Japanese in which reading annotations are placed above kanji; Ruby-OPD uses it as a representational-layer device, using orthographic perturbation so that the model is exposed to both surface written forms and target pronunciations during training.
50%UnverifiedRuby-OPD introduces On-Policy Distillation (OPD) to continue optimizing pronunciation accuracy on large-scale text-only corpora without paired speech.
50%UnverifiedOn the Joyo-Kanji-Yomi benchmark, Ruby-OPD achieves an accuracy of 97.13%, a Target Phoneme Error Rate (Target PER) of 2.02%, a Sentence-level Phoneme Error Rate (Sentence PER) of 0.70%, and a Sentence-level Character Error Rate (Sentence CER) of 4.21% under raw text input.
50%UnverifiedThe acquisition cost of Ruby annotations, the design details of the perturbation strategy, and the stability of OPD on larger-scale corpora remain engineering issues that need to be observed in follow-up work.
50%