Research map / Large language models / Language, safety and interpretability
Evaluating and Rewarding LALMs for Expressive Role-Play TTS via Mean Continuation Log-Probability
Yong Ren, Jingbei Li, Haiyang Sun, Yujie Chen, Cheng Yi, Yechang Huang, Hao Gu, Ye Bai, Xuerui Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SFT on WenetSpeech-RP-TTS to Step-Audio-2-mini (ablation) | Step-Audio-2-mini · Subjective | MOS: 1.856 -> 3.178 | Improved |
| add SFT (ablation) | Step-Audio-2-mini · W/ audio history | MCLP: -4.829 -> -4.725 | Improved |
| add SFT (ablation) | Step-Audio-2-mini · W/o audio history | CER (%): 12.007 -> 6.306 | Improved |
| add SFT (ablation) | Step-Audio-2-mini · W/ audio history | CER (%): 3.276 -> 3.334 | No clear effect |
| add GRPO RL with hybrid reward after SFT (ablation) | SFT model · Subjective | MOS: 3.178 -> 3.576 | Improved |
| add GRPO RL with hybrid reward after SFT (ablation) | SFT model · W/ audio history | MCLP: -4.725 -> -4.636 | Improved |
| add GRPO RL with hybrid reward after SFT (ablation) | SFT model · W/o audio history | MCLP: -4.731 -> -4.687 | Improved |
| add GRPO RL with hybrid reward after SFT (ablation) | SFT model · W/ audio history | CER (%): 3.334 -> 1.130 | Improved |
| add GRPO RL with hybrid reward after SFT (ablation) | SFT model · W/o audio history | CER (%): 6.306 -> 1.625 | Improved |
| add CER content reward to MCLP reward (ablation) | Proposed RL · W/ audio history | CER (%): 61.144 -> 1.130 | Improved |
| add CER content reward to MCLP reward (ablation) | Proposed RL · Subjective | MOS: 1.145 -> 3.576 | Improved |
| add CER content reward to MCLP reward (ablation) | Proposed RL · W/ audio history | MCLP: -4.590 -> -4.636 | Hurt |
Closest papers by meaning
- CharacterBench: Benchmarking Character Customization of Large Language Models (AAAI 2025 · similarity 0.86)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · similarity 0.85)
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.85)
- Codifying Character Logic in Role-Playing (NeurIPS 2025 · similarity 0.85)
- Aligning Spoken Dialogue Models from User Interactions (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
