You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
You agree that you will use the dataset solely for the purpose of JAPANESE COPYRIGHT ACT ARTICLE 30-4
Log in or Sign Up to review the conditions and access this dataset content.
J-CHAT is a Japanese large-scale dialogue speech corpus. For the detailed explanation, please see our paper
Disclaimer
TO USE THIS DATASET, YOU MUST AGREE THAT YOU WILL USE THE DATASET SOLELY FOR THE PURPOSE OF JAPANESE COPYRIGHT ACT ARTICLE 30-4.
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
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.
loading dataset
Please see the following to load the dataset as lhotse.CutSet
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)
For the info about lhotse.CutSet
please see the lhotse documentation
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
Contributors
- Wataru Nakata/中田 亘
- Kentaro Seki/関 健太郎
- Hitomi Yanaka/谷中 瞳
- Yuki Saito/齋藤 佑樹
- Shinnosuke Takamichi/高道 慎之介
- Hiroshi Saruwatari/猿渡 洋
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).
- Downloads last month
- 37