S2S-Arena Dataset

This repository hosts the S2S-Arena dataset. It covers four practical domains with 21 tasks, includes 154 instructions of varying difficulty levels, and features a mix of samples from TTS synthesis, human recordings, and existing audio datasets.

Introduction

GitHub Repository

For more information and access to the dataset, please visit the GitHub repository: S2S-Arena on GitHub

Related Publication

For detailed insights into the dataset’s construction, methodology, and applications, please refer to the accompanying academic publication: [coming soon]

Data Description

The dataset includes labeled audio files, textual emotion annotations, language translations, and task-specific metadata, supporting fine-grained analysis and application in machine learning. Each entry follows this format:

{
    "id": "emotion_audio_0",
    "input_path": "./emotion/audio_0.wav",
    "text": "[emotion: happy]Kids are talking by the door",
    "task": "Emotion recognition and expression",
    "task_description": "Can the model recognize emotions and provide appropriate responses based on different emotions?",
    "text_cn": "孩子们在门旁说话",
    "language": "English",
    "category": "Social Companionship",
    "level": "L3"
}

1. id: Unique identifier for each sample
2. input_path: Path to the audio file
3. text: English text with emotion annotation
4. task: Primary task associated with the data
5. task_description: Task description for model interpretability
6. text_cn: Chinese translation of the English text
7. language: Language of the input
8. category: Interaction context category
9. level: Difficulty or complexity level of the sample

"Some data also includes a noise attribute, indicating that noise has been added to the current sample and specifying the type of noise."

BIb

@misc{jiang2025s2sarenaevaluatingspeech2speechprotocols,
      title={S2S-Arena, Evaluating Speech2Speech Protocols on Instruction Following with Paralinguistic Information}, 
      author={Feng Jiang and Zhiyu Lin and Fan Bu and Yuhao Du and Benyou Wang and Haizhou Li},
      year={2025},
      eprint={2503.05085},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.05085}, 
}