Easy-Turn-Testset

# Easy Turn: Integrating Acoustic and Linguistic Modalities for Robust Turn-Taking in Full-Duplex Spoken Dialogue Systems

<p align="center">
Guojian Li¹, Chengyou Wang¹, Hongfei Xue¹,
Shuiyuan Wang¹, Dehui Gao¹, Zihan Zhang²,
Yuke Lin², Wenjie Li², Longshuai Xiao²,
Zhonghua Fu¹^,╀, Lei Xie¹^,╀
</p>

<p align="center">
¹ Audio, Speech and Language Processing Group (ASLP@NPU), Northwestern Polytechnical University <br>
² Huawei Technologies, China <br>
</p>

<div align="center">

| 🎤 [Demo Page](https://aslp-lab.github.io/Easy-Turn/) | 🤖 [Easy Turn Model](https://huggingface.co/ASLP-lab/Easy-Turn) | 📑 [Paper](https://arxiv.org) | 🌐 [Huggingface](https://huggingface.co/collections/ASLP-lab/easy-turn-68d3ed0b294df61214428ea7) |
|:---:|:---:|:---:|:---:|

</div>

<p align="center">
<img src="src/logo.png" alt="Institution 5" style="width: 600px; border-radius: 30px;">
</p>

## Download
The Easy Turn resources are available at [Model](https://huggingface.co/ASLP-lab/Easy-Turn), [Trainset](https://huggingface.co/datasets/ASLP-lab/Easy-Turn-Trainset), and [Testset](https://huggingface.co/datasets/ASLP-lab/Easy-Turn-Testset).

## Easy Turn Testset

In addition to the Easy Turn Trainset, we also release a speech test set—Easy Turn Testset, designed to evaluate turn-taking detection performance. It includes four dialogue turn states: 300 samples each for complete and incomplete, and 100 samples each for backchannel and wait. Real and synthetic speech are balanced at a 1:1 ratio. The transcriptions of Easy Turn Testset come from sources outside the training set, covering both casual conversations and human-computer interactions. Dialogue turn states are manually annotated to ensure higher accuracy. The Easy Turn Testset includes two types of speech: real recordings from internal speakers and synthetic speech generated with CosyVoice 2, using Emilia as the reference corpus and unseen speeches outside the training set as references. This design ensures the test set’s independence and diversity.

## Citation
Please cite our paper if you find this work useful:

# Easy Turn：融合声学与语言模态实现鲁棒全双工口语对话系统轮次管理

<p align="center">
李国建¹, 王承友¹, 薛鸿飞¹,
王水元¹, 高德辉¹, 张子涵²,
林宇科², 李文杰², 肖龙帅²,
付中华¹^,╀, 谢磊¹^,╀
</p>

<p align="center">
¹ 西北工业大学音频、语音与语言处理小组（ASLP@NPU）<br>
² 中国华为技术有限公司<br>
</p>

<div align="center">

| 🎤 [演示页面](https://aslp-lab.github.io/Easy-Turn/) | 🤖 [Easy Turn模型](https://huggingface.co/ASLP-lab/Easy-Turn) | 📑 [论文](https://arxiv.org) | 🌐 [Huggingface数据集集合](https://huggingface.co/collections/ASLP-lab/easy-turn-68d3ed0b294df61214428ea7) |
|:---:|:---:|:---:|:---:|

</div>

<p align="center">
<img src="src/logo.png" alt="机构标识" style="width: 600px; border-radius: 30px;">
</p>

## 下载
Easy Turn的相关资源可从以下链接获取：[模型](https://huggingface.co/ASLP-lab/Easy-Turn)、[训练集（Trainset）](https://huggingface.co/datasets/ASLP-lab/Easy-Turn-Trainset)与[测试集（Testset）](https://huggingface.co/datasets/ASLP-lab/Easy-Turn-Testset)。

## Easy Turn测试集

除Easy Turn训练集（Trainset）外，我们还发布了一款语音测试集——Easy Turn测试集，用于评估轮次检测性能。该测试集涵盖四类对话轮次状态：完整轮次与不完整轮次各300条样本，回馈语（backchannel）与等待轮次（wait）各100条样本。真实语音与合成语音的比例为1:1，实现均衡分布。

Easy Turn测试集的转录文本均来自训练集以外的数据源，涵盖日常会话与人机交互场景。对话轮次状态均经过人工标注，以确保更高的标注精度。

Easy Turn测试集包含两类语音：一是内部发言人的真实录制语音，二是基于CosyVoice 2生成的合成语音——该合成语音以Emilia作为参考语料库，并以训练集以外的未见过的语音作为参考样本。该设计保障了测试集的独立性与多样性。

## 引用说明
若您认为本研究对您的工作有所帮助，请引用我们的论文：

搜集汇总

背景与挑战

背景概述

Easy-Turn-Testset是一个用于评估对话系统中轮流检测性能的语音测试集，包含四种对话状态（完整、不完整、反馈和等待），样本数量分别为300、300、100和100，真实和合成语音比例为1:1。数据集的转录文本来自训练集之外的来源，覆盖了日常对话和人机交互场景，所有对话状态均经过人工标注。数据集包含真实录音和合成语音，后者使用CosyVoice 2生成，确保了测试集的独立性和多样性。