guilinhu/libri_conversation

---
license: mit
task_categories:
- audio-to-audio
language:
- en
---

# Libri Conversation FLAC Dataset

This dataset accompanies the paper:

**[Proactive Hearing Assistants that Isolate Egocentric Conversations](https://www.arxiv.org/abs/2511.11473)**
*Hu et al., 2025*

It contains **~234 hours of conversational-style audio** derived from LibriSpeech-like sources, processed into 60-second multispeaker segments under different experimental conditions:

- **libri_leaving** — scenarios where one speaker intermittently leaves the conversation
- **libri_multi** — 3-speaker conversational segments

All audio is stored in **FLAC** format. Metadata files (JSON) are preserved exactly as in the original directory structure.

---

## Dataset Structure

The dataset is organized into two main components:

libri_leaving/
train/
val/
test/

libri_multi/
train/
val/
test/

Because of Hugging Face API request limits, the dataset is packaged into `.tar` archives.
Each archive mirrors the original folder structure:

libri_leaving_train.tar
libri_leaving_val.tar
libri_leaving_test.tar

libri_multi_train.tar
libri_multi_val.tar
libri_multi_test.tar

---

## Citation

If you use this dataset, please cite the following paper:

@inproceedings{hu2025proactive,
title={Proactive Hearing Assistants that Isolate Egocentric Conversations},
author={Hu, Guilin and Itani, Malek and Chen, Tuochao and Gollakota, Shyamnath},
booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
pages={25377--25394},
year={2025}
}

许可证: MIT
任务类别：
- 音频到音频
语言：
- 英语

# Libri Conversation FLAC 数据集

本数据集配套论文：

**[《分离自我中心对话的主动听力助手》](https://www.arxiv.org/abs/2511.11473)**
*胡等人，2025年*

本数据集包含约234小时的会话类音频，数据源自类LibriSpeech数据源，经处理后生成不同实验条件下的60秒多说话人片段：

- **libri_leaving** — 单说话人间歇性退出对话的场景
- **libri_multi** — 三说话人会话片段

所有音频均存储为**FLAC格式**。元数据文件（JSON格式）严格遵循原始目录结构完整留存。

---

## 数据集结构

本数据集分为两大核心组件：

libri_leaving/
train/
val/
test/

libri_multi/
train/
val/
test/

由于Hugging Face API存在请求限制，本数据集被打包为.tar归档文件。每个归档文件均镜像原始文件夹结构：

libri_leaving_train.tar
libri_leaving_val.tar
libri_leaving_test.tar

libri_multi_train.tar
libri_multi_val.tar
libri_multi_test.tar

---

## 引用说明

若您使用本数据集，请引用以下论文：

@inproceedings{hu2025proactive,
title={Proactive Hearing Assistants that Isolate Egocentric Conversations},
author={Hu, Guilin and Itani, Malek and Chen, Tuochao and Gollakota, Shyamnath},
booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
pages={25377--25394},
year={2025}
}