Datasets:

Languages:
Japanese
Tags:
music
License:

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.

Please read README.md and LICENSE.md before downloading this corpus.

Log in or Sign Up to review the conditions and access this dataset content.

English version follows the Japanese text.

IdolSongsJp: アイドルグループ楽曲スタイルにもとづく音楽コーパス

日本のアイドルグループを模した 15 のオリジナル楽曲からなるコーパスです。

ティザームービー

ティザームービー

コーパス構成

楽曲と歌唱者

15 楽曲のうち、8 楽曲が女性グループ、7 曲が男性グループによる楽曲です。

女性歌唱者は 10 名、男性歌唱者は 8 名であり、歌唱メンバーは楽曲によって異なります。

作曲者・編曲者・作詞者はすべてアイドルグループに楽曲提供経験を持つプロフェッショナルです。各歌唱者はセミプロフェッショナルもしくはプロフェッショナルのボーカリストです(実際のアイドルではありません)。

楽曲の一覧と制作者を下記に示します。楽曲長は切り捨てです。BPM は代表値を示しています。

楽曲 ID 楽曲名 作詞 作曲 編曲 楽曲長 BPM 歌唱者数
f01-intro_juice いんとろじゅーす ハイジナカムラ ハイジナカムラ ハイジナカムラ 4:02 186 7
f02-aoharu_syndrome アオハルシンドローム ハイジナカムラ ハイジナカムラ ハイジナカムラ 4:27 190 6
f03-awake Awake あんたれす あんたれす あんたれす 4:28 175 5
f04-awanai_kagi 合わない鍵 Kashi Kashi 横山広稀 3:54 123 8
f05-kaerimichi 帰り道 Kashi Kashi 長谷部隼一 6:00 90 6
f06-ima_sekaiwa_kagayaiteru 今、世界は輝いてる 横井香菜&Kashi 横井香菜 長谷部隼一 4:54 132 4
f07-illuminyati イルミにゃティ 中野晃 中野晃 中野晃 3:19 160 7
f08-tic_tac_toe tic-tac-toe 横井香菜 横井香菜 SA!D 3:24 120 9
m01-liberty_wing Liberty wing Atohs Music Atohs Music Atohs Music 4:28 135 8
m02-trick_star trick star 長谷部隼一 長谷部隼一 長谷部隼一 4:34 134 7
m03-answer Answer でぃーの でぃーの Atohs Music 3:48 185 6
m04-uraomote_docchi 裏表どっち? 横山広稀 横山広稀 横山広稀 3:37 157 7
m05-deep_breath Deep breath Atohs Music Atohs Music Atohs Music 3:13 130 5
m06-soredemo_sukidayo それでも、好きだよ 横井香菜 横井香菜 横山広稀 / pf: 野村美和 5:26 78 4
m07-suki_no_aizu 「好き」の合図 あんたれす あんたれす あんたれす 4:19 138 6

ファイル構成

すべての音声信号は 48kHz / 32bit 浮動小数点フォーマットの RIFF WAV で収録されています。整数フォーマットで利用する場合には、FabFilter Pro-L 2iZotope Ozone などのソフトウェアを利用してディザ処理を行うことを推奨します。

音響信号はすべて同期されています。

master_48k32b_-7LUFS

-7LUFS でマスタリングされたマスター版楽曲信号です。

マスタリングの最終段で、Pro-L 2 の Modern プリセットで +2dB ゲインを上げています。

楽器音のバランス調整など、マスタリングはこの信号を基準に制作されています。最終的には ITU-R BS 1770-3 基準で -7.1LUFS 以上 -7.0LUFS 以下になるように調節されています。

すべての楽曲で、リードボーカルのみ削除したマイナスワンバージョンと、ボーカルトラックをすべて削除したインストゥルメンタルバージョンを収録しています。また、F01『いんとろじゅーす』のみコール・mix を含んだバージョンを収録しています。なお、これらは上述のマスター版と同一の設定でマスタリングを行っているため、最終的なラウドネスは -7LUFS には調整されていません。

master_48k32b_no-limit / master_48k32b_-9LUFS

master_48k32b_-7LUFS と比較して、no-limit 版は最終段の Pro-L 2 を適用していないもの、-9LUFS 版は最終段の Pro-L 2 のゲインを +0dB にしたものです。

solo_48k32b_*

マスター版のボーカルを各歌唱者のソロ歌唱に置き換えてマスタリングしたものです。すべての楽曲–歌唱者の組み合わせで収録しており、マスタリング条件ごとにのべ 95 トラック収録されています。

これらのデータを作る際は、ボーカルのトラックのみ各歌唱者ごとに個別に追加でミキシングしています。ハモり、コール、ボーカルを利用したエフェクト(『帰り道』の「粉雪」や『Deep breath』の「Weaving the Days」など)も含めて同一の歌唱者の歌声を利用しています。

ボーカルトラック以外のマスタリングの条件はマスター版と同一です。

stem_48k32b

楽器のカテゴリ別に書き出したステム信号です。カテゴリは MoisesDB に準じていますが、MoisesDB よりさらに細かな分類をしている場合があります。

マスタリングの際のエフェクトは適用されていません。マスター版の信号は、この信号を線形に加算したものにマスターエフェクトを適用して制作されています。

stem4_48k32b_-7LUFS / stem4_48k32b_-9LUFS / stem4_48k32b_no-limit

Demucs などの音源分離ソフトウェアでデファクトスタンダードになっている、bass / drums / other / vocal の 4 カテゴリ別のステム信号です。

これらは、stem_48k32b のうち該当するトラックを線形に加算したものにマスターエフェクトを適用することで制作されています。マスターエフェクトは非線形なため、この信号を足し合わせてもマスター版の信号とは一致しません。あくまで音量やステレオのバランスを近付けるために適用されています。

vocals_48k32b

ミックスの際に使用されている、エフェクトが適用されていないドライなボーカル信号です。メインメロディ、ハモり、コールなどトラック別に収録されています。制作の際には複数テイクを結合したのち、Melodyne を利用して手作業で修正しています。

chords / keys / lyrics / lyrics_kana

  • chords ディレクトリには、Harte の記法にしたがったコードのアノテーションが収録されています。すべての楽曲について 2 名以上の専門家がアノテーションを行い、そのコンセンサスをもとに整備されています。各列は開始秒、終了秒、コードに対応します。
  • keys ディレクトリには、転調を含む各楽曲の調性のアノテーションが含まれます。chords ディレクトリと同様の条件でアノテーションされています。
  • lyrics ディレクトリには歌詞のテキストファイルが含まれます。F01『いんとろじゅーす』のファイルにはコール・mix が含まれているバージョンを収録しています。
  • lyrics_kana ディレクトリには、漢字をひらがなに変換し、記号などを削除して正規化した歌詞のテキストファイルが含まれます。
    • 助詞の「は」や「へ」のみ「わ」や「え」に正規化しています。

コーパスの利用条件

このコーパスに含まれるすべての楽曲は産業技術総合研究所が著作権を保有しています。

このコーパスは、おおまかに次のルールにしたがって利用できます。詳細はライセンスファイルを確認してください。

  • 非商用の研究用途は無償利用可能
    • 利用の際には下記の論文を引用してください
    • プロダクトに組み込むなどの商用利用は個別にご相談ください
  • 非商用もしくは軽微な商用のエンターテインメントは無償利用可能
    • ここでいう軽微な商用とは、本コーパスの楽曲配信を主目的としていない、SNS および動画投稿サイトでの収益化、ライブコンサートでの演奏などを指します
    • 楽曲販売、音楽サブスクリプションサービスでの配信や、本コーパスの楽曲そのものを主たるコンテンツとした収益化付き動画の投稿などの商用利用は個別にご相談ください
    • 非商用の場合を含め、利用の際には、下記の論文もしくは「© 2025 産業技術総合研究所」の著作権表示が必要です
  • 各楽曲は、各制作者の著作者人格権を侵害しない範囲で、アレンジおよび二次創作可
    • カバー、編曲、替え歌などを想定しています
  • マスター音源や stem 信号に含まれる楽器音の信号について、機械学習による学習禁止・サンプリングなどによる別コンテンツ制作への利用禁止
    • これらの楽器音はあくまで作品・音楽コンテンツとして収録されており、ソフトウェア音源やサンプリング音源の目的外使用を懸念するためです
    • 音源分離技術やコード認識技術などにおける評価セットとしての利用は可能です
    • ボーカル信号は機械学習での利用が可能ですが、誹謗中傷をはじめとした不適切な音声の合成など人格・名誉を傷つける利用は禁止します
  • 宗教活動および政治活動での利用禁止
  • 楽曲制作者やボーカリストが持つ著作隣接権などの人権を侵害しうる利用禁止

引用

論文で記述する際は、以下の論文を引用してください。

Hitoshi Suda, Junya Koguchi, Shunsuke Yoshida, Tomohiko Nakamura, Satoru Fukayama, and Jun Ogata, “IdolSongsJp Corpus: A multi-singer song corpus in the style of Japanese idol groups,” in Proceedings of the 26th International Society for Music Information Retrieval (ISMIR) Conference, September 2025.

@INPROCEEDINGS{IdolSongsJp,
  title     = "{IdolSongsJp corpus: A multi-singer song corpus in the style of
               Japanese idol groups}",
  author    = "Suda, Hitoshi and Koguchi, Junya and Yoshida, Shunsuke and
               Nakamura, Tomohiko and Fukayama, Satoru and Ogata, Jun",
  booktitle = "{Proc. the 26th International Society for Music Information
               Retrieval (ISMIR) Conference}",
  year      =  2025
}

連絡先

産業技術総合研究所 須田仁志 [email protected]

IdolSongsJp: A Multi-Singer Japanese Idol-Style Song Corpus

IdolSongsJp contains 15 original tracks in the style of Japanese idol groups, each features multiple vocalists.

Corpus Overview

Songs and Singers

8 songs are performed by female groups and 7 by male groups. The corpus features 10 female singers and 8 male singers in total, and the combinations of singers vary from song to song.

All compositions, arrangements and lyrics were written by professionals who have provided songs for real-world idol groups. Every vocalist is a semi-professional or professional singer (not actual idols).

Song ID Title Length BPM # Singers
f01-intro_juice Introjuice 4:02 186 7
f02-aoharu_syndrome Aoharu Syndrome 4:27 190 6
f03-awake Awake 4:28 175 5
f04-awanai_kagi Awanai Kagi 3:54 123 8
f05-kaerimichi Kaerimichi 6:00 90 6
f06-ima_sekaiwa_kagayaiteru Ima, Sekai wa Kagayaiteru 4:54 132 4
f07-illuminyati Illuminyati 3:19 160 7
f08-tic_tac_toe tic-tac-toe 3:24 120 9
m01-liberty_wing Liberty wing 4:28 135 8
m02-trick_star trick star 4:34 134 7
m03-answer Answer 3:48 185 6
m04-uraomote_docchi Uraomote Docchi? 3:37 157 7
m05-deep_breath Deep breath 3:13 130 5
m06-soredemo_sukidayo Soredemo, Sukidayo 5:26 78 4
m07-suki_no_aizu “Suki” no Aizu 4:19 138 6

Directory Structure

All audio signals are provided as RIFF-WAV files at 48 kHz / 32-bit float and are time-aligned. If you need integer PCM, dithering with software such as FabFilter Pro-L 2 or iZotope Ozone is recommended.

master_48k32b_-7LUFS

These files are the mastered mixes with loudness normalized to -7 LUFS. The final limiter is FabFilter Pro-L 2 (Modern preset) with +2 dB gain. The total loudness is adjusted to between -7.1 and -7.0 LUFS according to ITU-R BS.1770-3.

For every song, this corpus also includes a minus-one mix (lead vocals muted) and an instrumental mix (all vocals muted). Only F01 “Introjuice” includes an additional “calls & mixes” version. These alternate mixes share the same processing chain and are not re-normalized to -7 LUFS.

master_48k32b_no-limit / master_48k32b_-9LUFS

Compared to master_48k32b_-7LUFS, _no-limit removes the final limiter entirely, and _-9LUFS keeps the limiter but sets its gain to 0 dB.

solo_48k32b_*

These folders contain versions in which the vocal stem has been replaced by a solo performance from each individual singer. All other processing was unchanged. All 95 possible song–singer combinations are included for every of 3 mastering conditions.

stem_48k32b

Instrument-category stems based on the MoisesDB taxonomy (sometimes split more finely). No mastering effects are applied to these stems. The master mix was created by summing these stems and then adding the mastering effect chain.

stem4_48k32b_*

Four-part stems (bass / drums / other / vocal) commonly used by music source separation tools such as Demucs. They were generated by summing the corresponding stem_48k32b tracks and applying the same mastering effects to master mixes. Because those effects are non-linear, adding the four stems does not reproduce the exact master. These effects are applied only to match volume and stereo balance to master mix.

vocals_48k32b

Dry, unprocessed vocal tracks used during mixing. Separate tracks are provided for lead melodies, harmonies and calls. Multiple takes were comped and manually tuned with Melodyne. This corpus includes solo vocal recordings in which each assigned singer performs the entire song, regardless of the original vocal part (song divison) assignments.

chords / keys / lyrics / lyrics_kana

  • chords: Chord annotations in Harte's notation, agreed upon by at least two experts. Each row lists start time, end time and chord label.
  • keys: Key annotations prepared under the same procedure to chords.
  • lyrics: Lyric text files. For F01, the calls & mixes version is also included.
  • lyrics_kana: Lyrics converted to hiragana and normalized by stripping of symbols.

Terms of Use

The National Institute of Advanced Industrial Science and Technology (AIST) owns the copyright to every track. The following list summarizes the license terms. For details, please refer to the license file.

  • You may use the corpus for non-commercial research free of charge. Any publication must cite the paper shown below.
    • Commercial use, including integration into products, requires prior approval by the copyright owner.
  • Entertainment use that is non-commercial or only lightly monetized is free.
    • Light monetization refers to slight monetization on social media/video platforms where the corpus is not the main content.
    • Selling the songs themselves or posting monetized videos that rely primarily on these tracks counts as commercial use and needs separate prior permission.
    • In all cases you must give proper credit, either by citing the paper or by adding “© 2025 National Institute of Advanced Industrial Science and Technology”.
  • You may create derivative works such as remixes or parody lyrics as long as you respect the authors' moral rights.
  • Using the instrumental signals for machine-learning training or as sampled material in new unrelated content is prohibited.
    • These tracks were produced for entertainment and are not intended to be used as training data for generative AI or related technologies.
    • You may use it for evaluation tasks such as source separation or chord recognition.
  • You may use the isolated vocal tracks for machine learning, but generating defamatory or otherwise inappropriate speech that violates their moral rights is prohibited.
  • Use for criminal, religious, or political purposes is prohibited.

Citation

When referring to the corpus, cite the following paper.

Hitoshi Suda, Junya Koguchi, Shunsuke Yoshida, Tomohiko Nakamura, Satoru Fukayama, and Jun Ogata, “IdolSongsJp Corpus: A multi-singer song corpus in the style of Japanese idol groups,” in Proceedings of the 26th International Society for Music Information Retrieval (ISMIR) Conference, September 2025.

@INPROCEEDINGS{IdolSongsJp,
  title     = "{IdolSongsJp corpus: A multi-singer song corpus in the style of
               Japanese idol groups}",
  author    = "Suda, Hitoshi and Koguchi, Junya and Yoshida, Shunsuke and
               Nakamura, Tomohiko and Fukayama, Satoru and Ogata, Jun",
  booktitle = "{Proc. the 26th International Society for Music Information
               Retrieval (ISMIR) Conference}",
  year      =  2025
}

Contact

Hitoshi Suda [email protected]

Downloads last month
17