WenetSpeech-Chuan

<h1 align="center">
WenetSpeech-Chuan: A Large-Scale Sichuanese Corpus With Rich Annotation For Dialectal Speech Processing
</h1>

<p align="center">
Yuhang Dai¹^,*, Ziyu Zhang¹^,*, Shuai Wang⁴^,5,
Longhao Li¹, Zhao Guo¹, Tianlun Zuo¹,
Shuiyuan Wang¹, Hongfei Xue¹, Chengyou Wang¹,
Qing Wang³, Xin Xu², Hui Bu², Jie Li³,
Jian Kang³, Binbin Zhang⁵, Lei Xie¹^,╀
</p>
<p align="center">
¹ Audio, Speech and Language Processing Group (ASLP@NPU), Northwestern Polytechnical University <br>
² Beijing AISHELL Technology Co., Ltd. <br>
³ Institute of Artificial Intelligence (TeleAI), China Telecom <br>
⁴ School of Intelligence Science and Technology, Nanjing University <br>
⁵ WeNet Open Source Community <br>
</p>

<p align="center">
📑 <a href="https://arxiv.org/abs/2509.18004">Paper</a> &nbsp&nbsp | &nbsp&nbsp
🐙 <a href="https://github.com/ASLP-lab/WenetSpeech-Chuan">GitHub</a> &nbsp&nbsp | &nbsp&nbsp
🤗 <a href="https://huggingface.co/collections/ASLP-lab/wenetspeech-chuan-68bade9d02bcb1faece65bda">HuggingFace</a>
<br>
🎤 <a href="https://aslp-lab.github.io/WenetSpeech-Chuan/">Demo Page</a> &nbsp&nbsp | &nbsp&nbsp
💬 <a href="https://github.com/ASLP-lab/WenetSpeech-Chuan?tab=readme-ov-file#contact">Contact Us</a>
</p>

<div align="center">
<img width="800px" src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/logo/WenetSpeech-Chuan-Logo.png?raw=true" />
</div>

## Dataset
### WenetSpeech-Chuan Overview

* Contains 10,000 hours of large-scale Chuan-Yu dialect speech corpus with rich annotations, the largest open-source resource for Chuan-Yu dialect speech research.</li>
* Stores metadata in a single JSON file, including audio path, duration, text confidence, speaker identity, SNR, DNSMOS, age, gender, and character-level timestamps. Additional metadata tags may be added in the future.</li>
* Covers ten domains: Short videos, Entertainment, Live streams, Documentary, Audiobook, Drama, Interview, News and others.</li>

<div align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/figs/domain.png?raw=true" width="300" style="display:inline-block; margin-right:10px;" />
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/figs/quality_distribution.jpg?raw=true" width="300" style="display:inline-block;" />
</div>

### Metadata Format
We store all audio metadata in a standardized JSON format, where the core fields include `utt_id` (unique identifier for each audio segment), `rover_result` (ROVER result of three ASR transcriptions), `confidence` (confidence score of text transcription), `jyutping_confidence` (confidence score of Cantonese pinyin transcriptions), and `duration` (audio duration); speaker attributes include `speaker_id`, `gender`, and `age`; audio quality assessment metrics include `sample_rate`, `DNSMOS`, and `SNR`; timestamp information includes `timestamp` (precisely recording segment boundaries with `start` and `end`); and extended metadata under the `meta_info` field includes `program` (program name), `region` (geographical information), `link` (original content link), and `domain` (domain classification).

#### 📂 Content Tree

```
WenetSpeech-Chuan
├── metadata.jsonl
│
├── audio_labels/
│ ├── wav_utt_id.jsonl
│ ├── wav_utt_id.jsonl
│ ├── ...
│ └── wav_utt_id.jsonl
│
├── .gitattributes
└── README.md
```
#### Data sample（CN）：

###### metadata.jsonl

```
{
"utt_id": 原始长音频id,
"wav_utt_id": 转化为wav后的长音频id,
"source_audio_path": 原始长音频路径,
"audio_labels": 转化后的长音频切分出的短音频标签文件路径,
"url": 原始长音频下载链接
}
```

###### audio_labels/wav_utt_id.jsonl：

```
{
"wav_utt_id_timestamp": 以 转化为wav后的长音频id_时间戳信息 作为切分后的短音频id (type: str),
"wav_utt_id_timestamp_path": 短音频数据路径 (type: str),
"audio_clip_id": 该段短音频在长音频中的切分顺序编号,
"timestamp": 时间戳信息,
"wvmos_score": wvmos分数，衡量音频片段质量 (type: float),
"text": 对应时间戳的音频片段的抄本 (type: str),
"text_punc": 带标点的抄本 (type: str),
"spk_num": 音频片段说话人个数，single/multi (type: str)
"confidence": 抄本置信度 (type: float),
"emotion": 说话人情感标签 (type: str，eg: 愤怒),
"age": 说话人年龄标签 (type: int范围, eg: 中年（36~59）),
"gender": 说话人性别标签 (type: str，eg: 男/女),
}
```

#### Data sample（EN）：

###### metadata.jsonl

```
{
"utt_id": Original long audio ID,
"wav_utt_id": Converted long audio ID after transforming to WAV format,
"source_audio_path": Path to the original long audio file,
"audio_labels": Path to the label file of short audio segments cut from the converted long audio,
"url": Download link for the original long audio
}
```

###### audio_labels/wav_utt_id.jsonl：

```
{
"wav_utt_id_timestamp": Short audio segment ID, composed of the converted long audio ID + timestamp information (type: str),
"wav_utt_id_timestamp_path": Path to the short audio data (type: str),
"audio_clip_id": Sequence number of this short segment within the long audio,
"timestamp": Timestamp information,
"wvmos_score": WVMOS score, measuring the quality of the audio segment (type: float),
"text": Transcript of the audio segment corresponding to the timestamp (type: str),
"text_punc": Transcript with punctuation (type: str),
"spk_num": Number of speakers in the audio segment, single/multi (type: str),
"confidence": Confidence score of the transcript (type: float),
"emotion": Speaker’s emotion label (type: str, e.g., anger),
"age": Speaker’s age label (type: int range, e.g., middle-aged (36–59)),
"gender": Speaker’s gender label (type: str, e.g., male/female)
}
```

### WenetSpeech Usage
You can obtain the original video source through the `link` field in the metadata file (`metadata.json`). Segment the audio according to the `timestamps` field to extract the corresponding record. For pre-processed audio data, please contact us using the information provided below.

## Contact
If you have any questions or would like to collaborate, feel free to reach out to our research team via email: [e-mail] or [e-mail].

You’re also welcome to join our WeChat group for technical discussions, updates, and — as mentioned above — access to pre-processed audio data.

<p align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/raw/main/src/figs/wechat_2.png" width="300" alt="WeChat Group QR Code"/>
<em>Scan to join our WeChat discussion group</em>
</p>

<p align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Yue/raw/main/figs/[e-mail]" width="300" alt="Official Account QR Code"/>
</p>

# WenetSpeech-Chuan：面向方言语音处理的大规模川渝方言标注语料库

**作者**：戴宇航¹^,*、张子宇¹^,*、王帅⁴^,5、李龙昊¹、郭钊¹、左天伦¹、王水原¹、薛鸿飞¹、王承友¹、王青³、徐鑫²、卜辉²、李杰³、康健³、张彬彬⁵、谢磊¹^,╀

¹ 西北工业大学音频、语音与语言处理组（ASLP@NPU）<br>
² 北京爱数智慧科技有限公司<br>
³ 中国电信人工智能研究院（TeleAI）<br>
⁴ 南京大学智能科学与技术学院<br>
⁵ WeNet开源社区

<p align="center">📑 <a href="https://arxiv.org/abs/2509.18004">论文</a> &nbsp&nbsp | &nbsp&nbsp 🐙 <a href="https://github.com/ASLP-lab/WenetSpeech-Chuan">GitHub仓库</a> &nbsp&nbsp | &nbsp&nbsp 🤗 <a href="https://huggingface.co/collections/ASLP-lab/wenetspeech-chuan-68bade9d02bcb1faece65bda">HuggingFace数据集集合</a><br>🎤 <a href="https://aslp-lab.github.io/WenetSpeech-Chuan/">演示页面</a> &nbsp&nbsp | &nbsp&nbsp 💬 <a href="https://github.com/ASLP-lab/WenetSpeech-Chuan?tab=readme-ov-file#contact">联系我们</a></p>

<div align="center">
<img width="800px" src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/logo/WenetSpeech-Chuan-Logo.png?raw=true" />
</div>

## 数据集
### WenetSpeech-Chuan 总览
* 包含10000小时的大规模川渝方言语音语料库，附带丰富标注信息，是目前开源的规模最大的川渝方言语音研究资源。
* 元数据以单一JSON文件存储，涵盖音频路径、时长、文本置信度、说话人身份、信噪比（SNR, Signal-to-Noise Ratio）、DNSMOS、年龄、性别以及字符级时间戳，未来或将新增更多元数据标签。
* 涵盖十大领域：短视频、娱乐、直播、纪录片、有声书、戏剧、访谈、新闻及其他领域。

<div align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/figs/domain.png?raw=true" width="300" style="display:inline-block; margin-right:10px;" />
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/blob/main/src/figs/quality_distribution.jpg?raw=true" width="300" style="display:inline-block;" />
</div>

### 元数据格式
我们将所有音频元数据以标准化JSON格式存储，核心字段包括：
- `utt_id`：每条音频片段的唯一标识符
- `rover_result`：三种自动语音识别（ASR, Automatic Speech Recognition）转录结果的ROVER融合结果
- `confidence`：文本转录置信度得分
- `jyutping_confidence`：拼音转录置信度得分
- `duration`：音频时长
- 说话人属性：`speaker_id`、`gender`（性别）与`age`（年龄）
- 音频质量评估指标：`sample_rate`（采样率）、`DNSMOS`与`SNR`（信噪比）
- 时间戳信息：`timestamp`，精确记录片段起止点，包含`start`（起始时刻）与`end`（结束时刻）
- `meta_info`字段下的扩展元数据包括：`program`（节目名称）、`region`（地理信息）、`link`（原始内容链接）与`domain`（领域分类）

#### 📂 目录结构

WenetSpeech-Chuan
├── metadata.jsonl
│
├── audio_labels/
│ ├── wav_utt_id.jsonl
│ ├── wav_utt_id.jsonl
│ ├── ...
│ └── wav_utt_id.jsonl
│
├── .gitattributes
└── README.md

#### 数据样本（中文）
##### metadata.jsonl

{
"utt_id": 原始长音频标识符,
"wav_utt_id": 转换为WAV格式后的长音频标识符,
"source_audio_path": 原始长音频文件路径,
"audio_labels": 从转换后的长音频中切分出的短音频片段的标签文件路径,
"url": 原始长音频下载链接
}

##### audio_labels/wav_utt_id.jsonl

{
"wav_utt_id_timestamp": 以转换后的长音频标识符+时间戳信息组成的短音频片段ID（字符串类型）,
"wav_utt_id_timestamp_path": 短音频数据路径（字符串类型）,
"audio_clip_id": 该短音频片段在长音频中的序列编号,
"timestamp": 时间戳信息,
"wvmos_score": 用于衡量音频片段质量的WVMOS得分（浮点型）,
"text": 对应时间戳的音频片段的转录文本（字符串类型）,
"text_punc": 带标点符号的转录文本（字符串类型）,
"spk_num": 音频片段中的说话人数量，取值为single（单说话人）或multi（多说话人）（字符串类型）,
"confidence": 转录文本的置信度得分（浮点型）,
"emotion": 说话人情感标签（字符串类型，例如：愤怒）,
"age": 说话人年龄标签（整数范围，例如：中年（36~59岁））,
"gender": 说话人性别标签（字符串类型，例如：男/女）
}

#### 数据样本（英文）
##### metadata.jsonl

{
"utt_id": Original long audio ID,
"wav_utt_id": Converted long audio ID after transforming to WAV format,
"source_audio_path": Path to the original long audio file,
"audio_labels": Path to the label file of short audio segments cut from the converted long audio,
"url": Download link for the original long audio
}

##### audio_labels/wav_utt_id.jsonl

{
"wav_utt_id_timestamp": Short audio segment ID, composed of the converted long audio ID + timestamp information (type: str),
"wav_utt_id_timestamp_path": Path to the short audio data (type: str),
"audio_clip_id": Sequence number of this short segment within the long audio,
"timestamp": Timestamp information,
"wvmos_score": WVMOS score, measuring the quality of the audio segment (type: float),
"text": Transcript of the audio segment corresponding to the timestamp (type: str),
"text_punc": Transcript with punctuation (type: str),
"spk_num": Number of speakers in the audio segment, single/multi (type: str),
"confidence": Confidence score of the transcript (type: float),
"emotion": Speaker’s emotion label (type: str, e.g., anger),
"age": Speaker’s age label (type: int range, e.g., middle-aged (36–59)),
"gender": Speaker’s gender label (type: str, e.g., male/female)
}

### WenetSpeech 使用方式
用户可通过元数据文件（`metadata.json`）中的`link`字段获取原始视频资源，依据`timestamps`字段对音频进行切分以提取对应片段。如需获取预处理后的音频数据，请通过下方提供的联系方式与我们取得联系。

## 联系方式
若您有任何疑问或合作意向，可通过邮箱`[e-mail]`或`[e-mail]`与我们的研究团队取得联系。
您也可加入我们的微信技术交流群，获取最新动态与相关讨论——如前文所述，该群也可用于获取预处理后的音频数据。

<p align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Chuan/raw/main/src/figs/wechat_2.png" width="300" alt="WeChat Group QR Code"/>
<em>扫码加入我们的微信交流群</em>
</p>

<p align="center">
<img src="https://github.com/ASLP-lab/WenetSpeech-Yue/raw/main/figs/[e-mail]" width="300" alt="Official Account QR Code"/>
<em>扫码关注官方公众号</em>
</p>

搜集汇总

背景与挑战

背景概述

WenetSpeech-Chuan是一个用于方言语音处理的大规模四川话（川渝方言）语料库，包含10,000小时语音数据，是目前最大的开源川渝方言资源。该数据集提供丰富注释，包括元数据如音频路径、时长、文本置信度、说话人身份、音频质量指标和字符级时间戳，并覆盖短视频、娱乐、直播等十个领域，支持多样化的语音研究。