Elamur Research⌕
☰
home / papers / 2026 / PED: Route-Decoupled Diagnostics for Persona Consistency in Spoken Agents

PED: Route-Decoupled Diagnostics for Persona Consistency in Spoken Agents

Paper · 2026 · Weihao Liu, Junrui Wei, Zhao Zhang et al.

Summary

Spoken role-playing agents often slip out of character over long conversations. Existing benchmarks usually give a single holistic "voice + character" score, which does not show whether the fault comes from the generated words or from their acoustic delivery. The authors propose PED (Persona–Emotion Decoupling), a diagnostic framework that treats text and speech as two separately measured routes. Off-the-shelf emotion projectors map both routes into a common seven-label affect space. This yields two measures, textual (TEC) and acoustic (AEC) emotion consistency. Each system–persona–route combination receives its own reference anchor, built from fresh single-turn responses. PED then tracks turn-by-turn alignment to that anchor through a scripted dialogue with baseline, stress, and recovery phases.

When PED was used to evaluate two architectures, a Speech LLM that operates fully end to end and a pipeline chaining an LLM with a TTS stage, it revealed four patterns:

  • persona separability depends on anchor overlap and varies across architectures;
  • the text route drifts under stress toward neutral-dominated outputs;
  • text and audio consistency are only loosely coupled;
  • in cascaded systems, an explicit style cue shapes audio-route structure.

The value of PED is that it turns opaque overall grades into turn-level signals that point to the failing component.

Key points

  • Measurement space and anchors. The shared space is a probability simplex over 7 labels (angry, disgust, fear, happy, neutral, sad, surprised). The labels are treated as instrument coordinates, not ground-truth emotions. Anchors average K=20 stateless single-turn samples per (system, persona, route).
  • Dialogue protocol. Each persona runs through a fixed 25-turn script: Baseline (turns 1–5), Stress (6–20), and Recovery (21–25).
  • Metrics. PED computes:
    • pairwise anchor cosine geometry;
    • nearest-anchor persona classification accuracy;
    • per-turn anchor similarity, summarized by phase;
    • dominant-label frequencies;
    • cross-persona convergence.
  • Setup. Two systems of roughly 3B parameters (end-to-end vs. cascaded) were tested with three Big Five-based personas (C, E, N).
  • Design implications.
    • End-to-end models need more distinct, controllable persona prosody, because their acoustic anchors nearly overlap.
    • In cascaded pipelines, the text route is the bottleneck under stress. Removing the style cue weakens persona-distinct AEC structure.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Xi'an Jiaotong UniversityXi'an · China22 works · 36,046 citations🎓Tianjin UniversityTianjin · China41 works · 1,128 citations

Also on the byline: Tianjin University of Technology.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Weihao Liu, Junrui Wei, Zhao Zhang and Ju Zhang. 2026. PED: Route-Decoupled Diagnostics for Persona Consistency in Spoken Agents. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.445

@inproceedings{liu2026ped,
  title = {{PED: Route-Decoupled Diagnostics for Persona Consistency in Spoken Agents}},
  author = {Weihao Liu, Junrui Wei, Zhao Zhang and Ju Zhang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.445},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.445}
}

References

19 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Marwa Abdulhai, Ryan Cheng, Donovan Clay, Tim Althoff, Sergey Levine, and Natasha Jaques. 2025. Consistently simulating human personas with multi-turn reinforcement learning. Preprint, arXiv:2511.00222. — [arXiv:2511.00222](https://arxiv.org/abs/2511.00222) find
  2. Jingyi Chen, Zhimeng Guo, Jiyun Chun, Pichao Wang, Andrew Perrault, and Micha Elsner. 2025. Do audio LLMs really LISTEN, or just transcribe? measuring lexical vs. acoustic emotion cues reliance. Preprint, arXiv:2510.10444. find
  3. Pedro Corrêa, João Lima, Victor Moreno, Lucas Ueda, and Paula Dornhofer Paro Costa. 2025. Evaluating emotion recognition in spoken language models on emotionally incongruent speech. Preprint, arXiv:2510.25054. — [arXiv:2510.25054](https://arxiv.org/abs/2510.25054) find
  4. Jr. Costa, Paul T. and Robert R. McCrae. 1992. Revised NEO Personality Inventory (NEO-PI-R) and NEO Five-Factor Inventory (NEO-FFI) Professional Manual. Psychological Assessment Resources, Odessa, FL. find
  5. Yassine El Boudouri, Walter Nuninger, Julian Alvarez, and Yvan Peter. 2025. Role-playing evaluation for large language models. Preprint, arXiv:2505.13157. — [arXiv:2505.13157](https://arxiv.org/abs/2505.13157) find
  6. Changhao Jiang, Jiajun Sun, Yifei Cao, Jiabao Zhuang, Hui Li, Baoyu Fan, Tao Ji, Tao Gui, and Qi Zhang. 2025. SpeechRole: A large-scale dataset and benchmark for evaluating speech role-playing agents. Preprint, arXiv:2508.02013. — [arXiv:2508.02013](https://arxiv.org/abs/2508.02013) find
  7. Wenyu Li, Xiaoqi Jiao, Yi Chang, Guangyan Zhang, and Yiwen Guo. 2025. AudioRole: An audio dataset for character role-playing in large language models. Preprint, arXiv:2509.23435. — [arXiv:2509.23435](https://arxiv.org/abs/2509.23435) find
  8. Junru Lu, Jiazheng Li, Guodong Shen, Lin Gui, Siyu An, Yulan He, Di Yin, and Xing Sun. 2025. RoleMRC: A fine-grained composite benchmark for role-playing and instruction-following. In Findings of the Association for Computational Linguistics: ACL 2025, pages 21008–21030, Vienna, Austria. Association… find
  9. Hosein Mohebbi, Grzegorz Chrupała, Willem Zuidema, Afra Alishahi, and Ivan Titov. 2024. Disentangling textual and acoustic features of neural speech representations. Preprint, arXiv:2410.03037. — [arXiv:2410.03037](https://arxiv.org/abs/2410.03037) find
  10. Qwen Team. 2024. Qwen2.5: A party of foundation models. Blog post. find
  11. Vinay Samuel, Henry Peng Zou, Yue Zhou, Shreyas Chaudhari, Ashwin Kalyan, Tanmay Rajpurohit, Ameet Deshpande, Karthik Narasimhan, and Vishvak Murahari. 2025. PersonaGym: Evaluating persona agents and LLMs. Preprint, arXiv:2407.18416. find
  12. Jiatong Shi, Jionghao Han, Yichen Lu, Santiago Pascual, Pengfei Wu, Chenye Cui, Shinji Watanabe, Chao Weng, and Cong Zhou. 2025. Speech-DRAME: A framework for human-aligned benchmarks in speech role-play. Preprint, arXiv:2511.01261. — [arXiv:2511.01261](https://arxiv.org/abs/2511.01261) find
  13. Rongfeng Su, Changqing Xu, Xinyi Wu, Feng Xu, Xie Chen, Lan Wang, and Nan Yan. 2024. Investigating acoustic-textual emotional inconsistency information for automatic depression detection. Preprint, arXiv:2412.18614. — [arXiv:2412.18614](https://arxiv.org/abs/2412.18614) find
  14. Quan Tu, Shilong Fan, Zihang Tian, and Rui Yan. 2024. CharacterEval: A chinese benchmark for role-playing conversational agent evaluation. Preprint, arXiv:2401.01275.
  15. Weihao Wu, Liang Cao, Xinyu Wu, Zhiwei Lin, Rui Niu, Jingbei Li, and Zhiyong Wu. 2025. VoxRole: A comprehensive benchmark for evaluating speech-based role-playing agents. Preprint, arXiv:2509.03940. — [arXiv:2509.03940](https://arxiv.org/abs/2509.03940) find
  16. Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. 2025. Qwen2.5-omni technical report.
  17. Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. 2023. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. Preprint, arXiv:2305.11000.
  18. Haonan Zhang, Run Luo, Xiong Liu, Yuchuan Wu, Ting-En Lin, Pengpeng Zeng, Qiang Qu, Feiteng Fang, Min Yang, Lianli Gao, Jingkuan Song, Fei Huang, and Yongbin Li. 2025. OmniCharacter: Towards immersive role-playing agents with seamless speech-language personality interaction. Preprint, arXiv:2505.202… find
  19. Siyi Zhou, Yiquan Zhou, Yi He, Xun Zhou, Jinchao Wang, Wei Deng, and Jingchen Shu. 2025. Indextts2: A breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech. Preprint, arXiv:2506.21619. — [arXiv:2506.21619](https://arxiv.org/abs/2506.21619) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/liu-2026-ped-route-decoupled-diagnostics-persona/