Transfer of Speech-Rewarded Style Planning Across TTS Backends

Evaluate whether Speech-Rewarded Style Planning transfers effectively to text-to-speech backends other than CosyVoice3, given that both reward computation and synthesis in the reported experiments use CosyVoice3.

Background

Speech-Rewarded Style Planning is trained using teacher-forced target-speech likelihoods computed by a frozen Fun-CosyVoice3-0.5B model, and the same CosyVoice3 backend is used for synthesis during evaluation. Consequently, the reported gains establish effectiveness only within that specific TTS backend and do not determine whether the learned style-planning behavior generalizes to other synthesizers.

The conclusion explicitly identifies cross-backend transfer as unresolved future work, making this a concrete open evaluation problem concerning the robustness and portability of the proposed method.