BAF: an audio fingerprinting dataset for broadcast monitoring

<strong>Overview</strong> Broadcast Audio Fingerprinting dataset is an open, available upon request, annotated dataset for the task of music monitoring in broadcast. It contains 2,000 tracks from Epidemic Sound's private catalogue as reference tracks that represent 74 hours. As queries, it contains over 57 hours of TV broadcast audio from 23 countries and 203 channels distributed with 3,425 one-min audio excerpts. It has been annotated by six annotators in total and each query has been cross-annotated by three of them obtaining high inter-annotator agreement percentages, which validates the annotation methodology and ensures the reliability of the annotations. <strong>Purpose of the dataset</strong> This dataset aims to become the standard dataset to evaluate Audio Fingerprinting algorithms since it’s built on real data, without the use of any data-augmentation techniques. It is also the first dataset to address background music fingerprinting, which is a real problem in royalties distribution. <strong>Dataset use</strong> This dataset is available for conducting non-commercial research related to audio analysis. It shall not be used for music generation or music synthesis. <strong>About the data</strong> All audio files are monophonic, 8kHz, 128kb/s, pcm_s16le encoded in .wav. Annotations mark which tracks sound (either in foreground or background) in each query (if any) and also the specific times where it starts and ends sound in the query. <em>Note that there are 88 queries that do not have any matches.</em> For more information check the dedicated Github repository: https://github.com/guillemcortes/baf-dataset and the dataset datasheet included in the files. <strong>Dataset contents</strong> The dataset is structured following this schema <pre><code>baf-dataset/ ├── baf_datasheet.pdf ├── annotations.csv ├── changelog.md ├── cross_annotations.csv ├── queries_info.csv ├── queries │ ├── query_0001.wav │ ├── query_0002.wav │ ├── … │ └── query_3425.wav ├── queries_info.csv └── references ├── ref_0001.wav ├── ref_0002.wav ├── … └── ref_2000.wav</code></pre> There are two folders named <strong>queries</strong> and <strong>references</strong> containing the <em>wav</em> files of TV broadcast recordings and the reference tracks, respectively. <strong>annotations.csv</strong> file contains the annotations made by the 6 annotators, giving the following information: annotations.csv content summary query reference query_start query_end annotator query_0692.wav ref_1235.wav 0.0 59.904 annotator_6 <strong>cross_annotations.csv </strong>contains the resulting annotations after merging the overlapping annotations in <strong>annotations.csv</strong> file. x_tag has three different values: single: the segment has only been annotated by one annotator. majority: the segment has been annotated by two annotators. unanimity: the segment has been annotated by the three annotators. cross_annotations.csv content summary query reference query_Start query_end annotators x_tag query_0693.wav ref_1834.wav 37.53 38.07 ['annotator_3'] single query_0693.wav ref_1834.wav 18.18 37.48 ['annotator_3', 'annotator_5', 'annotator_3'] unanimity query_0693.wav ref_1834.wav 37.48 37.53 ['annotator_5', 'annotator_3'] majority <strong>queries_info.csv </strong>contains information about the queries as a citation reference. It contains the country, the channel and the date where the broadcast happened. queries_info.csv content summary filename country channel datetime query_0001.wav Norway Discovery Channel 2021-02-26 14:45:26 <strong>changelog.md </strong>contains a curated, chronologically ordered list of notable changes for each version of the dataset. <strong>baf_datasheet.pdf</strong> contains standardized documentation for datasets <strong>Ownership of the data</strong> Next, we specify the ownership of all the data included in BAF: Broadcast Audio Fingerprinting dataset. For licensing information, please refer to the “License” section. <strong>Reference tracks</strong> The reference tracks are owned by Epidemic Sound AB, which has given a worldwide, revocable, non-exclusive, royalty-free licence to use and reproduce this data collection consisting of 2,000 low-quality monophonic 8kHz downsampled audio recordings. <strong>Query tracks</strong> The query tracks come from publicly available TV broadcast emissions so the ownership of each recording belongs to the channel that emitted the content. We publish them under the right of quotation provided by the Berne Convention. <strong>Annotations</strong> Guillem Cortès together with Alex Ciurana and Emilio Molina from BMAT Music Licensing S.L. have managed the annotation therefore the annotations belong to BMAT. <strong>Accessing the dataset</strong> The dataset is available upon request. Please include, in the justification field, your academic affiliation (if you have one) and a brief description of your research topics and why you would like to use this dataset. Bear in mind that this information is important for the evaluation of every access request. <strong>License</strong> This dataset is available for conducting non-commercial research related to audio analysis. <strong>It shall not be used for music generation or music synthesis</strong>. Given the different ownership of the elements of the dataset, the dataset is licensed under the <strong>following conditions</strong>: User’s access request Research only, non-commercial purposes No adaptations nor derivative works Attribution to Epidemic Sound and the authors as it is indicated in the ”citation” section. <strong>Please include, in the justification field, your academic affiliation (if you have one) and a brief description of your research topics and why you would like to use this dataset.</strong> <strong>Acknowledgments</strong> With the support of Ministerio de Ciencia Innovación y universidades through Retos-Colaboración call, reference: RTC2019-007248-7, and also with the support of the Industrial Doctorates Plan of the Secretariat of Universities and Research of the Department of Business and Knowledge of the Generalitat de Catalunya. Reference: DI46-2020.<strong>概述</strong> 广播音频指纹（Broadcast Audio Fingerprinting）数据集是一款开源、需申请获取的标注数据集，用于广播电视场景下的音乐监测任务。该数据集包含来自Epidemic Sound私人曲库的2000条参考曲目，总时长达74小时。作为查询集，其收录了来自23个国家、203个电视频道的超过57小时的电视广播音频，以3425条1分钟时长的音频片段形式分发。本数据集总计由6名标注人员完成标注，每条查询音频均由其中3名标注人员开展交叉标注，获得了较高的标注者间一致性（inter-annotator agreement）百分比，这一结果验证了标注方法的有效性，同时确保了标注结果的可靠性。 <strong>数据集用途</strong> 本数据集旨在成为评估音频指纹（Audio Fingerprinting）算法的标准数据集，因其基于真实数据构建，未使用任何数据增强技术。这也是首个针对背景音乐指纹识别任务的公开数据集，而该任务是版权费分配领域的实际痛点问题。 <strong>数据集使用规范</strong> 本数据集仅可用于与音频分析相关的非商业研究，不得用于音乐生成或音乐合成任务。 <strong>数据说明</strong> 所有音频文件均为单声道、8kHz采样率、128kb/s，采用pcm_s16le编码格式封装为.wav文件。标注信息会标记每条查询音频中以前景或背景形式出现的参考曲目（若存在），以及该曲目在查询音频中的起止时间。<em>请注意，共有88条查询音频未匹配到任何参考曲目。</em> 如需了解更多信息，请查看专属GitHub仓库：https://github.com/guillemcortes/baf-dataset 以及文件中附带的数据集说明书。 <strong>数据集内容</strong> 数据集按照以下结构组织： <pre><code>baf-dataset/ ├── baf_datasheet.pdf ├── annotations.csv ├── changelog.md ├── cross_annotations.csv ├── queries_info.csv ├── queries │ ├── query_0001.wav │ ├── query_0002.wav │ ├── … │ └── query_3425.wav ├── queries_info.csv └── references ├── ref_0001.wav ├── ref_0002.wav ├── … └── ref_2000.wav</code></pre> 数据集包含两个分别名为<strong>queries</strong>和<strong>references</strong>的文件夹，分别存放电视广播录音音频文件与参考曲目音频文件。<strong>annotations.csv</strong>文件包含6名标注人员的标注结果，包含以下字段：query、reference、query_start、query_end、annotator。示例如下：query_0692.wav, ref_1235.wav, 0.0, 59.904, annotator_6。 <strong>cross_annotations.csv</strong> 包含合并了<strong>annotations.csv</strong>中重叠标注后的最终标注结果。x_tag包含三种不同取值：single：该片段仅由1名标注人员标注；majority：该片段由2名标注人员标注；unanimity：该片段由全部3名标注人员标注。cross_annotations.csv包含以下字段：query、reference、query_start、query_end、annotators、x_tag。示例如下： query_0693.wav, ref_1834.wav, 37.53, 38.07, ['annotator_3'], single query_0693.wav, ref_1834.wav, 18.18, 37.48, ['annotator_3', 'annotator_5', 'annotator_3'], unanimity query_0693.wav, ref_1834.wav, 37.48, 37.53, ['annotator_5', 'annotator_3'], majority <strong>queries_info.csv</strong> 包含查询集的相关信息，可作为引用参考。该文件包含广播发生的国家、电视频道以及日期时间。queries_info.csv包含以下字段：filename、country、channel、datetime。示例如下：query_0001.wav, Norway, Discovery Channel, 2021-02-26 14:45:26。 <strong>changelog.md</strong> 包含按时间顺序整理的数据集各版本的重要变更记录。<strong>baf_datasheet.pdf</strong> 包含数据集的标准化说明文档。 <strong>数据所有权</strong> 下文将明确BAF（广播音频指纹）数据集包含的所有数据的所有权归属。关于授权许可信息，请参阅“许可协议”章节。 <strong>参考曲目</strong> 参考曲目归Epidemic Sound AB所有，该机构已授予全球范围内、可撤销、非排他、免版权费的使用与复制该数据集的许可，该数据集包含2000条经过8kHz降采样的低质量单声道音频录音。 <strong>查询音频</strong> 查询音频来自公开可获取的电视广播信号，因此每条录音的所有权归播出该内容的电视频道所有。我们依据《伯尔尼公约》规定的引用权发布此类音频。 <strong>标注内容</strong> Guillem Cortès与来自BMAT Music Licensing S.L.的Alex Ciurana及Emilio Molina共同完成了标注工作，因此标注内容归BMAT所有。 <strong>数据集获取方式</strong> 本数据集需申请方可获取。请在申请理由栏中注明您的学术机构（若有）、研究主题的简要描述，以及您希望使用该数据集的原因。请注意，此类信息将用于评估每一项访问申请。 <strong>许可协议</strong> 本数据集仅可用于与音频分析相关的非商业研究，不得用于音乐生成或音乐合成任务。鉴于数据集各组成部分的所有权归属不同，本数据集需遵循以下许可条件：仅用于学术研究、非商业用途；不得进行修改或创作衍生作品；需按照“引用”章节的要求注明对Epidemic Sound及原作者的归属。请在申请理由栏中注明您的学术机构（若有）、研究主题的简要描述，以及您希望使用该数据集的原因。 <strong>致谢</strong> 本项目得到了西班牙科学创新与大学部通过Retos-Colaboración计划的支持，项目编号：RTC2019-007248-7；同时得到了加泰罗尼亚自治区工商与知识部大学与研究秘书处的工业博士计划支持，项目编号：DI46-2020。