---
license: other
language:
- ja
size_categories:
- 1M<n<10M
extra_gated_prompt: "You agree that you will use the dataset solely for the purpose of JAPANESE COPYRIGHT ACT ARTICLE 30-4"
---

J-CHAT is a Japanese large-scale dialogue speech corpus.
For the detailed explanation, please see our [paper](https://arxiv.org/abs/2407.15828)

# PLEASE READ THIS FIRST
>[!IMPORTANT]
> TO USE THIS DATASET, YOU MUST AGREE THAT YOU WILL USE THE DATASET SOLELY FOR THE PURPOSE OF JAPANESE COPYRIGHT ACT ARTICLE 30-4.

# What's new?
>[!NOTE] 
> Add transcription of corpus. transcripition are based on [reazonspeech-nemo-v2](https://huggingface.co/reazon-research/reazonspeech-nemo-v2)


# How can I use this data for commercial purposes?
Commercial use is not admitted. ~~If you want to use this data for commercial purposes, Please build one by yourself.
Corpus construction programs are distributed on [Github](https://github.com/sarulab-speech/J-CHAT)~~

The Github repo isn't ready yet. But we will be releasing the code soon. Stay tuned!!

# How to use
## Requirements
The dataet loading will require following python libraries.
* [lhotse](https://github.com/lhotse-speech/lhotse)
* [webdataset](https://github.com/webdataset/webdataset)
* If you need transcription, [smart-open](https://github.com/getcrest/smart-open)
## loading dataset
Please see the following to load the dataset as `lhotse.CutSet`

### Without transcription
```python
import lhotse

# change the following line to the subset and the data domain you like.
# For data domain the options are youtube and podcast.
# For subset, train, valid, test, others are available.
# For instance, if you want to get data from youtube test set. the filepath would be filelists/yotube_test.txt
with open("filelists/podcast_train.txt") as f:
    urls = f.read().splitlines()

cutset = lhotse.CutSet.from_webdataset(urls)
```

### With transcription
```python
import json
import lhotse

with open("transcribed_jchat/podcast_train.json") as f:
  fields = json.load(f)

cuts = lhotse.CutSet.from_shar(fields=fields)

```

For the info about `lhotse.CutSet` please see the [lhotse documentation](https://lhotse.readthedocs.io/en/latest/)

# LICENSE
CC-BY-NC 4.0

TO USE THIS DATASET, YOU MUST AGREE THAT YOU WILL USE THE DATASET SOLELY FOR THE PURPOSE OF JAPANESE COPYRIGHT ACT ARTICLE 30-4.

# Contact
We have ensured that our dataset does not infringe on any rights of the original data holders. 
However, if you wish to request the removal of your data from the dataset, please feel free to contact us at the email address below:

shinnosuke_takamichi [*at*] keio.jp

# Other resources
* [Speech samples generated with dGSLM trained on J-CHAT](https://sarulab-speech.github.io/j-chat/dgslm_speech_sample/)
* [dGSLM model weights](https://github.com/sarulab-speech/j-chat/tree/master/weights)

# Contributors
* [Wataru Nakata/中田 亘](https://wataru-nakata.github.io)
* [Kentaro Seki/関 健太郎](https://trgkpc.github.io/)
* [Hitomi Yanaka/谷中 瞳](https://hitomiyanaka.mystrikingly.com/)
* [Yuki Saito/齋藤 佑樹](https://sython.org/)
* [Shinnosuke Takamichi/高道 慎之介](https://sites.google.com/site/shinnosuketakamichi/home)
* [Hiroshi Saruwatari/猿渡 洋](https://researchmap.jp/read0102891)

# 謝辞/acknowledgements
本研究は、国立研究開発法人産業技術総合研究所事業の令和5年度覚醒プロジェクトの助成を受けたものです。
/This work was supported by AIST KAKUSEI project (FY2023).