# StyleSet

**A spoken language benchmark for evaluating speaking-style-realted speech generation**  
Released in our paper, [Audio-Aware Large Language Models as Judges for Speaking Styles](https://arxiv.org/abs/2506.05984)


![image/png](https://cdn-uploads.huggingface.co/production/uploads/622326ae0129f2097d69a3e2/Q8Os1g5vfy22Y9myvSc7X.png)

---

## Tasks

1. **Voice Style Instruction Following**  
   - Reproduce a given sentence verbatim.  
   - Match specified prosodic styles (emotion, volume, pace, emphasis, pitch, non-verbal cues).

2. **Role Playing**  
   - Continue a two-turn dialogue prompt in character.  
   - Generate the next utterance with appropriate prosody and style.
   - The dataset is modified from IEMOCAP with the consent of the authors. Please refer to [IEMOCAP](https://sail.usc.edu/iemocap/) for details and the original data of IEMOCAP. We do not redistribute the data here.

---

## Evaluation

We use ALLM-as-a-judge for evaluation. Currently, we found that `gemini-2.5-pro-0506` reaches the best agreement with human evaluators.
The complete evaluation prompt and evaluation pipelines can be found in Table 3 to Table 5 in our paper.


## Citation

If you use StyleSet or find ALLM-as-a-judge useful, please cite our paper by
```
@misc{chiang2025audioawarelargelanguagemodels,
      title={Audio-Aware Large Language Models as Judges for Speaking Styles}, 
      author={Cheng-Han Chiang and Xiaofei Wang and Chung-Ching Lin and Kevin Lin and Linjie Li and Radu Kopetz and Yao Qian and Zhendong Wang and Zhengyuan Yang and Hung-yi Lee and Lijuan Wang},
      year={2025},
      eprint={2506.05984},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2506.05984}, 
}
```