HPRO: Hierarchical Progressive Reward Optimization via Preference Extraction for Emotional Text-to-Speech
HPRO is a framework that improves emotional expression in Text-to-Speech systems by solving conflicts between content and emotion optimization, and sparse rewards and dense generation. It uses a differentiable reward model called HD-Emo to separate style from content, while progressively aligning objectives across frame, word, and sentence levels to achieve better emotional quality without losing speech intelligibility.