Broader validation across model scales and TTS architectures

Validate the kana-domain ASR reward for reinforcement-learning post-training of Japanese text-to-speech systems across broader model scales and TTS architectures, beyond the single 0.5-billion-parameter base model evaluated in the study.

Background

The study evaluates the proposed Kana-CER reinforcement-learning reward using one 0.5-billion-parameter Sarashina2.2-TTS base model. Although the reward substantially reduces target-kanji reading errors while preserving comparable orthographic CER and speech-quality measures, the results do not establish whether these gains generalize to models of different sizes or to other TTS architectures.

The authors explicitly identify broader validation as future work, making cross-scale and cross-architecture evaluation the unresolved problem. Such validation would determine the robustness and transferability of pronunciation-oriented kana-domain ASR rewards beyond the experimental configuration used in this paper.