ISMIR04 Genre Identification task dataset

This is a collection of audio used for the Genre Identification task of the ISMIR 2004 audio description contest organized by the Music Technology Group (Universitat Pompeu Fabra). The audio for the task was collected from Magnatune, which contains a large amount of music licensed under Creative Commons licenses. The task of the contest was to classify a set of songs into genres, using the genre labels that Magnatune provided in their database. Further information about the original contest and the contents of the dataset can be obtained from the following technical report: Cano P, Gómez E, Gouyon F, Herrera P, Koppenberger M, Ong B, Serra X, Streich S, Wack N. ISMIR 2004 audio description contest. Barcelona: Universitat Pompeu Fabra, Music technology Group; 2006. 20 p. Report No.: MTG-TR-2006-02 http://hdl.handle.net/10230/34013 The original contest website can be found at http://ismir2004.ismir.net/genre_contest/ The dataset contains the audio tracks from following 8 genres: classical, electronic, jazz- & blues, metal-, punk, rock-, pop, world. For the genre recognition contest, the data was grouped into 6 classes: classical, electronic, jazz-blues, metal-punk, rock-pop, world, where in some cases two genres were merged into a single class. Note that ground-truth files uses these 6 classes, however in some cases the data is organised by original genre. Audio The audio is in MP3 format. It is divided into three folders, representing different subsets of the collection. Each folder has 729 files, split into classes. The number of files in each category reflects the proportion of files in each category in Magnatune when the dataset was created. No track appears in more than one folder. Training: files for generating a classification model, arranged by class. Development: A separate set of files for participants to test their model against. Evaluation: originally a private subset, the files used to evaluate the accuracy of all submitted models The training and development set each consist of: classical: 320 files electronic: 115 files jazz_blues: 26 files metal_punk: 45 files rock_pop: 101 files world: 122 files The evaluation set consists of 729 tracks with a similar distribution. Metadata Each folder of audio has a corresponding folder containing metadata of the files in that folder. The metadata is included in a file, tracklist.csv which has the following headers: class, artist, album, track, track number, file path The evaluation tracklist file has an additional column representing the magnatune track id of the recording. Due to the way that the data was collected and distributed for the challenge, the metadata for the development subset is anonymised. Licensing The audio is licensed under a CC Attribution-NonCommercial-ShareAlike license (https://creativecommons.org/licenses/by-nc-sa/1.0/). Using this dataset We would highly appreciate if scientific publications of works partly based on this dataset cite the above publication. We are interested in knowing if you find our datasets useful! If you use our dataset please email us at [e-mail] and tell us about your research.

本数据集为音乐流派识别任务专用音频集合，源自西班牙庞培法布拉大学（Universitat Pompeu Fabra）音乐技术组（Music Technology Group）主办的2004年国际音乐信息检索大会（ISMIR 2004）音频描述竞赛。

本次竞赛所用音频均采集自Magnatune平台，该平台拥有大量采用知识共享（Creative Commons）许可协议授权的音乐作品。本次竞赛的核心任务为：基于Magnatune数据库提供的流派标签，将指定歌曲集按音乐流派进行分类。

关于本次原始竞赛及数据集内容的详细信息，可参考以下技术报告：Cano P, Gómez E, Gouyon F, Herrera P, Koppenberger M, Ong B, Serra X, Streich S, Wack N. ISMIR 2004音频描述竞赛. 巴塞罗那：庞培法布拉大学音乐技术组；2006. 共20页，报告编号：MTG-TR-2006-02，访问链接：http://hdl.handle.net/10230/34013。竞赛原始官网地址为：http://ismir2004.ismir.net/genre_contest/。

本数据集包含以下8个音乐流派的音频曲目：古典、电子、爵士与布鲁斯、金属、朋克、摇滚、流行、世界音乐。本次流派识别竞赛中，主办方将数据合并为6个类别：古典、电子、爵士-布鲁斯、金属-朋克、摇滚-流行、世界音乐，部分原流派被合并为单一类别。需注意，真值标签文件采用上述6个类别，但部分数据仍按原始流派进行组织。

音频格式方面，所有音频均为MP3格式。数据集分为三个文件夹，对应不同的子集划分，每个文件夹均包含729个音频文件，并按类别分组。每个类别的文件数量反映了数据集构建时Magnatune平台上对应类别的作品占比。所有曲目仅在一个文件夹中出现，无跨文件夹重复的曲目。

三个子集的具体说明如下：
1. 训练集：用于生成分类模型的音频文件，按类别组织。
2. 开发集：供参赛选手测试模型性能的独立数据集。
3. 评估集：原属非公开子集，用于对所有提交的模型进行精度评估。

训练集与开发集的类别分布如下：
- 古典：320个文件
- 电子：115个文件
- 爵士-布鲁斯：26个文件
- 金属-朋克：45个文件
- 摇滚-流行：101个文件
- 世界音乐：122个文件

评估集共包含729条曲目，类别分布与整体数据集近似。

元数据部分：每个音频文件夹均对应一个包含对应文件元数据的文件夹，元数据存储于tracklist.csv文件中，表头包含："class"（类别）、artist（艺术家）、album（专辑）、track（曲目名）、track number（曲目编号）、file path（文件路径）。评估集的tracklist.csv额外包含一列Magnatune曲目ID。由于本次挑战赛的数据采集与分发方式，开发集的元数据已做匿名化处理。

许可协议：本数据集的音频采用知识共享署名-非商业性使用-相同方式共享许可协议（https://creativecommons.org/licenses/by-nc-sa/1.0/）。

数据集使用说明：若基于本数据集开展相关研究并发表学术成果，恳请引用上述技术报告。我们十分期待了解您对本数据集的使用反馈！若您使用本数据集，请发送邮件至[e-mail]告知您的研究工作。