D-LPJ Decouples Multi-Dimensional Prosody Scoring in TTS Evaluation
September 18, 2026
Researchers introduce Decoupled-Live-ProsodyJudge to prevent verdict coupling in TTS evaluation, where judges align all dimension scores with a single preference. The method distills Gemini capabilities into Qwen3-Omni to provide cost-effective, fine-grained prosody metrics for emotion, intonation, and energy.
HOW THIS AFFECTS YOU
●
builderYou can use this distilled model for low-cost reinforcement learning feedback in speech synthesis pipelines.
●
researcherYou can avoid evaluation bias where judges collapse multi-dimensional rubrics into single preference bits.