mls_sidon

# MLS-Sidon

## Overview
This dataset is a **cleansed version of Multilingual LibriSpeech (MLS)** with **Sidon** speech restoration mode for **Speech Synthesis** and **Spoken Language Modeling**.

The dataset is provided in **[WebDataset](https://github.com/webdataset/webdataset) format** for efficient large-scale training.

- **Source**: [Multilingual LibriSpeech](https://www.openslr.org/94/)
- **Languages**: English, German, French, Spanish, Italian, Polish, Dutch, Portuguese
- **Format**: WebDataset (`.tar` shards)
- **License**: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)

---

## Dataset Structure
Each sample in the dataset contains:

- **`flac`** — audio file (48 kHz, single channel)
- **`metadata.json`** *(optional)* — metadata including language, speaker ID, and original MLS reference

Example (inside a `.tar` shard):
```

000001.flac
000001.metadata.json
000002.flac
000002.metadata.json
...

````

---

## How to Use

### With 🤗 Datasets
You can load the WebDataset directly with Hugging Face’s `datasets` library:

```python
import datasets
from IPython.display import Audio
from huggingface_hub import hf_hub_download
import yaml

base_url = "https://huggingface.co/datasets/sarulab-speech/mls_sidon/resolve/main/"
language = 'english'
split = 'test'
data_file_path = hf_hub_download(repo_id="sarulab-speech/mls_sidon", repo_type="dataset", filename="paths.yaml")
paths = yaml.load(open(data_file_path, "r"), Loader=yaml.FullLoader)

ds = datasets.load_dataset("webdataset", data_files=[base_url + p for p in paths['english'][split]],streaming=True)['train']

sample = next(iter(ds))
audio = sample['flac']
print(sample['metadata.json'])
Audio(audio['array'], rate=audio['sampling_rate'])
````

Replace `language` with the language (e.g., `english`, `german`).

---

## Citation

If you use this dataset, please cite Sidon and the original MLS paper:

```
@misc{nakata2025sidonfastrobustopensource,
title={Sidon: Fast and Robust Open-Source Multilingual Speech Restoration for Large-scale Dataset Cleansing},
author={Wataru Nakata and Yuki Saito and Yota Ueda and Hiroshi Saruwatari},
year={2025},
eprint={2509.17052},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2509.17052},
}
```

```
@inproceedings{pratap2020mls,
title = {MLS: A Large-Scale Multilingual Dataset for Speech Research},
author = {Pratap, Vineel and Xu, Qiantong and Sriram, Anuroop and others},
booktitle = {Interspeech},
year = {2020}
}
```

---

## License

This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).

---

## Acknowledgements

* **Original data**: [Multilingual LibriSpeech (MLS)](https://www.openslr.org/94/)

# MLS-Sidon

## 概述
本数据集是**多语言LibriSpeech（Multilingual LibriSpeech, MLS）**的净化版本，搭载了适用于**语音合成（Speech Synthesis）**与**口语语言建模（Spoken Language Modeling）**的**Sidon**语音修复模式。

本数据集采用**WebDataset**格式进行存储，以支持高效的大规模训练。

- **来源**：[多语言LibriSpeech（Multilingual LibriSpeech, MLS）](https://www.openslr.org/94/)
- **支持语言**：英语、德语、法语、西班牙语、意大利语、波兰语、荷兰语、葡萄牙语
- **数据格式**：WebDataset（`.tar`分卷文件）
- **授权协议**：[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)

---

## 数据集结构
数据集的每条样本包含以下内容：

- **`flac`** — 音频文件（采样率48 kHz，单声道）
- **`metadata.json`**（可选）—— 元数据，包含语言、说话人ID以及原始MLS引用信息

示例（`.tar`分卷文件内结构）：

000001.flac
000001.metadata.json
000002.flac
000002.metadata.json
...

---

## 使用方法
### 使用🤗 Datasets库
您可以直接通过Hugging Face的`datasets`库加载该WebDataset数据集：
python
import datasets
from IPython.display import Audio
from huggingface_hub import hf_hub_download
import yaml

base_url = "https://huggingface.co/datasets/sarulab-speech/mls_sidon/resolve/main/"
language = 'english'
split = 'test'
data_file_path = hf_hub_download(repo_id="sarulab-speech/mls_sidon", repo_type="dataset", filename="paths.yaml")
paths = yaml.load(open(data_file_path, "r"), Loader=yaml.FullLoader)

ds = datasets.load_dataset("webdataset", data_files=[base_url + p for p in paths['english'][split]], streaming=True)['train']

sample = next(iter(ds))
audio = sample['flac']
print(sample['metadata.json'])
Audio(audio['array'], rate=audio['sampling_rate'])

将`language`替换为对应语言（例如`english`、`german`）。

---

## 引用说明
若您使用本数据集，请引用Sidon以及原始MLS相关论文：
bibtex
@misc{nakata2025sidonfastrobustopensource,
title={Sidon: Fast and Robust Open-Source Multilingual Speech Restoration for Large-scale Dataset Cleansing},
author={Wataru Nakata and Yuki Saito and Yota Ueda and Hiroshi Saruwatari},
year={2025},
eprint={2509.17052},
archivePrefix={arXiv},
primaryClass={cs.SD},
url={https://arxiv.org/abs/2509.17052},
}

bibtex
@inproceedings{pratap2020mls,
title = {MLS: A Large-Scale Multilingual Dataset for Speech Research},
author = {Pratap, Vineel and Xu, Qiantong and Sriram, Anuroop and others},
booktitle = {Interspeech},
year = {2020}
}

---

## 授权协议
本数据集采用[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)协议发布。

---

## 致谢
* **原始数据来源**：[多语言LibriSpeech（MLS）](https://www.openslr.org/94/)