Abstract

The responsible management of multimodal music datasets plays a crucial role in the development and evaluation of music processing systems. However, navigating the landscape of legal and ethical considerations can be a complex and challenging task due to the magnitude and diversity of such. This paper clarifies these divergent legal and ethical considerations and highlights the challenges associated with multimodality and AI systems. Focusing on the most crucial stages of multimodal music data management, we provide recommendations for tackling legal and ethical challenges. We emphasize the importance of establishing an inclusive and accessible music data environment, encouraging researchers and data users to adopt responsible approaches towards managing multimodal music data collections.

Author Keywords

Multimodal, Music Processing, Datasets, Accessibility, Ethics, Legal, AI

Introduction

Multimodal music datasets incorporate different media and data types, such as MIDI, audio, video, and motion capture. The combination of multiple modalities has introduced various challenges and complexities, and careful decisions must be made to ensure a responsible data management process. Data owners and users facing the challenges of data management can find available solutions to assist them in data management planning1 and plan evaluation [1]. However, the abundance and scarcity of information in this area often contribute to confusion, particularly when trying to distinguish between what data managers must do and what they should do. One good example of helping researchers by clarifying the requirements of data management is the appropriate classification of legal information and sensitive data by the University of Oslo2. This system categorizes and color-codes data into “open (green)”, “restricted (yellow)”, “in confidence (red)”, and “strictly in confidence (black)”, providing clear guidelines for handling data based on its sensitivity and access restrictions. Such classifications not only aid in access control but also help users understand their responsibilities in protecting confidential information. Nonetheless, comprehensive systems of this nature are limited in capabilities, and availability and are not yet widespread or standardized.

In this paper, we aim to answer the following research questions:

• How does multimodality challenge music data management?

• What are the differences between legal and ethical frameworks in multimodal music dataset management?

• Which stages of the dataset lifecycle are most critical for adhering to legal and ethical frameworks, and why?

These questions collectively address the overarching problem of ensuring responsible and effective management of multimodal music datasets. By answering them, we seek to provide a comprehensive understanding of the intricate requirements, as well as the challenges presented by multimodality. It is important to note that many of the challenges discussed in this paper are not exclusive to the use of datasets for AI in music; however, we explore them specifically within the context of AI music. We explore the norms and gaps in related work, identify potential areas for improvement, and shed light on the differences between legal and ethical considerations that need to be taken into account, aiming to promote responsible practices of managing multimodal music datasets for the benefit of researchers, users, and the wider community.

Defining the Problem

As with many other disciplines, Artificial Intelligence (AI) is becoming increasingly important in music research, performance, and creation. At the same time, the music data collected to support music processing tasks has transformed from unimodal to multimodal, combining various sensory modalities captured using different media and data types, such as MIDI, audio, video, motion information, physiological measurements, and more. Multimodal music datasets encompass a variety of complementary information, which is crucial for training and evaluating AI models in music processing-related tasks. They enable researchers to explore complex interactions between different modalities, leading to advancements in areas such as music generation [2], classification [3], and recommendation systems [4].

Data management entails collecting, curating, and organizing data, from managing their raw version to publishing the final analysis results [5]. Here we focus on music data and specifically multimodal data used in AI systems. In contrast to unimodal music datasets, the management of multiple data types entails numerous challenges. For example, significant resources are required due to format disparities and data size [6]. To provide a clearer overview of the basic data management stages and corresponding challenges, we present the decisions that need to be made throughout the entire process. We highlight the importance of adhering to legal and ethical requirements, significantly influencing the resulting quality of the dataset and its management (Figure 1). For instance, a lack of diversity in the dataset may lead to unwanted bias, unclear documentation may risk transparency, and non-accessible data may lead to irreproducible research.

The lack of clear data management documentation is making the promotion of transparency challenging in the study and application of music [6]. This lack of transparency is not necessarily intentional, but rather a result of data managers being overwhelmed by the vast amount of information and legal frameworks available for data management and publishing. To address this, it would be helpful to provide a clear overview of the legal and ethical responsibilities carried by both dataset owners and users. Additionally, implementing FAIR guidelines (Findable, Accessible, Interoperable, and Reusable) and adopting best (or good enough) [7] practices can help ensure compliance with these guidelines. This paper discusses the importance of responsible data management, prioritizing inclusivity and accessibility when developing multimodal music datasets for AI systems.

By inclusivity, we mean that datasets represent diverse genres, languages, and cultural contexts, offering a wide range of information. This can significantly decrease unwanted bias in music processing tasks [8]. To ensure accessibility, it is important to consider how the information can reach a wider audience. Many datasets are currently released in a way that makes them primarily accessible to other experts. So, even if they are openly available and adhere to the FAIR guidelines (Findable, Accessible, Interoperable, and Reusable), they may not be accessible to, for example, students, non-peers, and non-experts. This entails ensuring that the dataset is approachable and comprehensible to users with varying levels of musical knowledge and expertise. Clear and user-friendly metadata, annotations, and explanations should be provided to aid understanding and navigation.

Legal Issues of Multimodal Music Datasets

Navigating the legal frameworks surrounding music data is challenging for most researchers. The plethora of terms, laws, and rules of legal obligations in distributing and using music data can be overwhelming. We see a need to provide clear data management guidelines to researchers. This section aims to clarify the legal prerequisites for handling music data within the appropriate legal frameworks and provide guidelines to assist creators and users of multimodal music datasets.

Copyrights

Intellectual Property Rights (IPR) encompass a body of laws designed to safeguard the interests of intellectual property owners, granting them exclusive rights of use [9]. The primary legal mechanisms governing the dissemination of ideas include patents (which provide exclusive rights for unique inventions), copyrights (which protect original artistic or literary works), and trademarks (which identify commercial origins) [9]. In this context, our attention is directed toward copyrights due to their specific relevance to music and our intention to not deal with inventions or commercial branding, but rather with the protection of artistic works.

Issues related to copyrights and music go way back in time, so one may wonder why it is still relevant to discuss this matter. Music advancements go hand-in-hand with technological breakthroughs, making it pivotal to copyright law development in creative endeavors. Throughout music’s evolution, music data has changed from physical scores and tape recordings to digital scores and digital audio recordings. These media also necessitate adding relevant metadata [10]. This implies that as technology evolves, the laws that protect the copyrights of music evolve with it. Or at least they should.

Music datasets usually consist of original artworks, whether that is a combination of dance performances [11], instrument performances [12], or a combination of pop songs and metadata [13]. Furthermore, we refer to music datasets closely connected to music processing tasks. Even though not much attention has been paid to the use and copyright of AI algorithms themselves [14], there has been a big shift in industry and research from valuing the algorithms of music creation to valuing the data being used [10]. The construction, distribution, and usage of multimodal music datasets introduce intricate legal and copyright challenges that demand careful consideration.

In some cases, multimodal music datasets may contain royalty-free music, which does not require any certain license [13]. There is the special case of AI-generated music data, which may not be subjected to copyrights, since "machines are not eligible for copyright assessment", or because in AI systems, the decisions are done "automatically" [15]. It is important to mention that even in such cases, the engineers who program the AI music models are responsible for the outcome of such data. An exception to this is the black box approach, which, in any case, does not fall under the scope of FAIR research due to the lack of explainability [16].

Managing copyrights for each unique modality can be complex, requiring different approaches for distinct data types. For example, using MIDI data to train an AI system usually requires fewer liability issues than audio files, even though both the composition and the recording are subjected to copyright protection [10]. One particular challenge in multimodal music data copyright arises in the use and distribution of audiovisual material. The legal frameworks governing audio and video content differ from those applicable to combined audiovisual material, highlighting the complexity of copyright considerations in multimedia contexts [17]. For instance, while an audio recording might be cleared for use, the accompanying video might have separate restrictions, leading to contradicting legal terms for audio versus audiovisual content.

The concept of "fair use," which allows limited usage without explicit permission for certain purposes like criticism, teaching, or research, adds complexity to the legal landscape. It is important to note that "fair use" is a concept that is specifically present in the legal framework of the United States. This concept presents particular challenges when dealing with datasets for training AI systems, as there is no specific legislation protecting such data. Consequently, these processes often fall under the umbrella of fair use. The recent introduction of the AI Act in Europe [18] aims to address these concerns within the European Union.

Licensing

Copyrights form the legal foundation for protecting original work. The terms for legal use and distribution of this copyrighted content are defined by licensing and it is one of the key legal requirements of multimodal music data management. This involves acquiring the necessary rights from artists, creators, and other rights holders. Several frameworks exist to facilitate licensing of music datasets, such as Creative Commons [19], or MIT licenses [20]. Proper attribution is essential to uphold licensing agreements and ensure responsible data management practices [21].

Licensing multimodal music datasets presents several challenges. For example, clearances for one modality, such as audio, may not automatically cover other modalities like images or text in the dataset. However, it is essential to obtain licensing agreements that cover all the necessary elements of the dataset, including the various modalities, to ensure legal compliance and avoid potential infringement issues. This can be achieved by either using one license for the whole dataset [22], or the combination of different licenses for one publication [23].

The European "right to be forgotten" presents a significant challenge in this case. This right allows individuals to request the removal of their personal data from public records, including datasets [24]. If an AI system has already been trained on such a dataset, it is unclear how to handle the removal request without compromising the integrity of the AI model. This issue has not yet been appropriately addressed in the legal or technological domains, adding another layer of complexity to the management of AI-generated music data.

Privacy

Multimodal music datasets usually contain personal information, like physiological measurements [25] or users’ profile information [26]. The protection of this data and the handling of it securely and confidentially belongs to the category of data privacy. When it comes to multimodal music datasets used in AI, data privacy becomes a crucial aspect to consider, especially when the datasets involve research participants taking part in music emotion recognition experiments [27] or indigenous music research [28].

Privacy regulations vary across countries [6]. In Europe, the General Data Protection Regulation (GDPR) provides guidelines for data collection, storage, processing, and transfer, imposing strict obligations on organizations handling personal data. The GDPR is also applied in the US, alongside state-specific regulations like the California Consumer Privacy Act (CCPA) that grant residents various rights regarding their personal information [29]. China has recently introduced the Personal Information Protection Law (PIPL), emphasizing informed consent and prohibiting unauthorized handling of personal data [30]. However, AI music platforms are still not regulated regarding the type of personal data they can use, and they can document every single interaction between a platform and its users [31]. The new European regulation for AI (AI Act) will soon enforce a new set of laws for AI companies to limit user profiling and privacy violations [18].

Different data types are protected by different privacy-related legal frameworks. For example, video recordings require heightened copyright protection due to the inclusion of sensitive information, prompting researchers to provide the video only upon request or publish the data after cropping/blurring out the identifying parts of the participants [32]. Combining personal and sensitive data in a single collection exponentially increases the risk of identification, making it challenging to ensure the privacy of participants. Anonymization is often considered a potential solution to address privacy concerns in multimodal music datasets [22]. However, it is important to note that AI advancements have made it increasingly possible to de-anonymize data [33][34][35], raising doubts about the effectiveness of traditional anonymization methods in safeguarding privacy.

In the context of research participants, data privacy considerations go beyond legal issues and encompass ethical concerns. It involves addressing questions regarding the ownership and co-creation of the data, as well as ensuring participants’ rights to privacy and informed consent.

Ethical Issues of Multimodal Music Datasets

Instead of relying solely on legal protection, it is the responsibility of data owners and users to establish an ethical framework for data usage (Table 1). Ethical frameworks, closely aligned with the principles of FAIR (Findable, Accessible, Interoperable, Reusable), promote data reuse independent of specific ethical regulations [36]. FAIR principles enhance the discoverability, accessibility, interoperability, and reusability of data by providing guidelines for data management practices, metadata standards, and infrastructure. In parallel, Open Data emphasizes unrestricted availability and sharing of data without limitations on access [37]. The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems further proposes five ethical principles for music AI systems: protection of human rights, prioritization of well-being, accountability of designers and operators, transparency, and minimal risk of misuse [38]. By combining legal compliance, ethical frameworks, FAIR principles, and Open Data concepts, responsible and ethical use of multimodal music datasets in AI can be fostered. In this section, we clarify the ethical considerations that contribute to the responsible use of multimodal music datasets for AI systems.

Table 1: Overview of the differences between legal and ethical frameworks in data management.

Acknowledgment of Original Work

As advancements in AI continue to revolutionize the way we interact with and create music, the ethical considerations surrounding data usage and content creation take on added significance. While legal frameworks provide essential guidelines for protecting intellectual property and data privacy, it is equally important to acknowledge and adhere to the ethical responsibilities inherent in utilizing artists’ images, biographical information, and derivative works. This transition from legal compliance to ethical engagement underscores the necessity of recognizing and respecting the rights of content creators and data subjects in the rapidly evolving landscape of AI-driven music processing.

It is customary in academia to acknowledge and build upon the work of other researchers by citing their publications. Similarly, the same principle should apply when utilizing academic outputs such as methods, source code, and data. Determining fair use is contingent on the specific context, and the concept of transformative use, which involves significantly modifying or repurposing content, may impact its assessment. Navigating the complexities of fair use, the introduction of the AI Act, and the absence of specific legislation protecting datasets used for AI training underscore the importance of recognizing and respecting the rights of content creators.

Each data type within a multimodal music dataset presents unique challenges regarding management techniques and adherence to ethical principles. For instance, music scores are collected in various formats and annotated by different performers. At the same time, MIDI, audio, video, and motion capture recordings originate from different devices with diverse compression methods and formats. Sensor data are usually captured in binary format, while questionnaires and interviews are extensively customizable and personalized across researchers [7]. Dealing with the ethical management of these diverse data types in a single music collection is undeniably challenging.

Tackling Unwanted Bias

Ethical considerations in multimodal music dataset creation are inherently complex compared to traditional datasets. Firstly, they go beyond copyright issues to encompass the responsible use and representation of diverse cultural and musical expressions. Concerns may arise regarding the appropriation, representation, and respectful treatment of music from different cultural contexts or communities, necessitating a nuanced approach to dataset construction and utilization. Including stereotypical opinions in datasets or models may exacerbate discriminatory assumptions [31]. Transparent AI, defined by the absence of discrimination against a specific group in the data used by the AI algorithm, is an important principle in this context [39].

There is a difference between fairness and bias. Fairness is related to the ability to explain how an AI model operates in a particular way or has arrived at a given result [15]. It is closely related to transparency and explainability, and it is something that not only contributes to ethical data management but also assists the work of the AI engineers by contributing to the clarification of metrics, less discrimination, and better overall team composition [15]. Conversely, biases refer to systematic favoritism in decision-making or data analysis that deviates from true or objective values. Bias can be introduced at different stages of the AI system, including data collection, labeling, feature selection, model training, or decision-making. Multimodal datasets have the potential to perpetuate biases present in the source data, creating the need for measures to mitigate bias, ensure equal representation, and accurately portray cultural nuances in music.

One notable example of bias in AI-based is seen in music recommendation systems. Studies have pointed out that these systems may be skewed toward maximizing consumer engagement and profitability rather than providing the personalized recommendations they claim to deliver. In [31], Born raises an important question: does the bias in recommendation systems stem exclusively from the dataset itself, or is it also influenced by the procedures used in dataset manipulation and model training? It helps to examine the same problem in another task, like Audio-Video Question Answering (AVQA). According to [8], employing techniques like data balancing can effectively reduce bias. This indicates that carefully selecting and processing data may help address bias issues in AI-based music systems. Handling explicit content responsibly, upholding ethical curation standards, and involving practitioners of living musical traditions in model training discussions are crucial [40]. One question remains whether any AI developer can freely use and exploit materials in the public domain, including traditional music, that are not protected by copyright [41].

Informed Consent

Ensuring privacy protection is not only a legal imperative but also an fundamental in scientific development. Informed consent stands as a critical ethical consideration for safeguarding privacy. It is pertinent to the types of licenses and data uses applied to the collected information. Participants must be fully apprised of the intended use of their data and granted the opportunity to explicitly consent to the storage and processing of their personal information for AI research purposes. Addressing these challenges, researchers should engage in clear communication with participants regarding the nature of data collection, the specific licenses utilized, the intended use of the data, the methods employed for anonymization, and the potential risks associated with re-identification. By providing such information, researchers foster transparency and promote participants’ informed consent. This approach upholds ethical principles while ensuring the protection of individuals’ privacy rights in scientific endeavors.

Multimodal music datasets often intersect multiple disciplines, such as musicology, computer science, data science, and psychology, necessitating comprehensive approaches that account for the distinct requirements of each discipline. Maintaining ethical integrity and complying with data protection regulations are essential when collecting user-generated data, such as music preferences or physiological responses. Prioritizing informed consent and anonymization of personal information is crucial in this process [22]. However, it is worth noting that while informed consent is not always a legal requirement for research, it remains a crucial aspect. One particular challenge in this regard is “dynamic consent”, which means that participants in a study should be able to revoke their consent at any time. That is easier said than done when datasets are shared openly and used as part of machine learning models [7].

Key Lifecycle Stages for FAIR Principles in Multimodal Music Datasets

Efficiently managing music datasets within AI systems presents a range of challenges at various stages of the dataset’s lifecycle. Figure 2 provides an overview of these challenges and crucial steps in effectively handling music data. From initial acquisition to processing, analysis, and eventual publishing or archiving, careful attention must be given to copyright compliance, property rights clearance, and storage infrastructure. Ethical and legal acquisition of music data, respecting privacy and intellectual property rights, is paramount. Following FAIR principles is crucial in data collection, which includes acquiring the necessary permissions and licenses for data use and distribution. These principles focus on ensuring datasets are accessible to all, rather than emphasizing the inclusivity of their content. Although inclusive data collection is ideally desirable, it is often impractical due to constraints like time, cost, and occasionally legal limitations. Publishing music datasets must also respect restrictions and attribution requirements. By addressing these challenges, practitioners and researchers can minimize legal risks and adopt responsible practices, enabling the development of AI applications in music that align with ethical considerations and legal regulations.

Data Collection

The initial stage of collecting (or acquiring) multimodal music data is crucial for any music processing task. Appropriate documentation and metadata facilitate the findability and reproducibility of the dataset for potential users [42], as researchers or practitioners can understand and utilize the data effectively. Clear metadata, including details such as data source, format, and context, enhances the accessibility of the dataset, allowing for easier discovery and utilization [13].

Data collection forms the foundation of any AI system. Ensuring that the data collected is accurate, representative, and free from bias is crucial. FAIR principles demand that the data used in AI systems accurately reflect the diverse range of individuals and populations it aims to serve. Failure to collect diverse and comprehensive data could result in biased models perpetuating discrimination, exclusion, or marginalization. Robust guidelines and protocols should be in place to detect and address biases in the data or the data collection process. Collecting data encompassing diverse perspectives and experiences helps prevent the amplification of biases during model training and deployment stages.

As mentioned earlier, adhering to ethical principles requires obtaining informed consent from individuals collecting data. Respecting privacy and data protection rights is crucial to maintain trust and ensure ethical practices. Consent is vital in ensuring that individuals have the necessary information and agency to decide how their data is used. Both legal and ethical consent are important considerations in data collection and use. Legally, obtaining consent ensures compliance with privacy laws and regulations, protecting individuals’ rights and privacy. Ethically, consent reflects respect for individuals’ autonomy, allowing them to make informed choices about participating in data collection and having the ability to withdraw their consent if desired. Without proper consent, data collection may infringe upon individuals’ privacy rights and may not align with ethical principles of fairness and respect.

Ensuring proper data acquisition involves respecting intellectual property rights and abiding by copyright laws. It is crucial to obtain the necessary permissions and provide proper attribution when using copyrighted data to avoid legal repercussions. Ethical sourcing practices should be followed, and data should not be acquired through unlawful or unethical means. Researchers must consider the origin and legitimacy of data sources, ensuring compliance with ethical guidelines and regulations. These practices contribute to maintaining integrity, transparency, and legal compliance in AI-driven endeavors.

Data Archiving and Publishing

Archiving and publishing are two distinct processes that play important roles in making research datasets accessible to the scientific community. Archiving involves the preservation and long-term storage of datasets, ensuring their retention and availability for future reference. It focuses on safeguarding the integrity of the dataset over time and is typically undertaken for historical and reference purposes. On the other hand, publishing entails actively disseminating datasets through established platforms, repositories, or journals to make them openly available to researchers and the broader scientific community. It aims to promote transparency, collaboration, and reproducibility of research findings by enabling immediate accessibility and sharing of datasets. Both archiving and publishing align with the purposes of ethical data management by facilitating the findability of data, enabling its reuse, and supporting the transparency of research outcomes.

Publishing and archiving make datasets findable by the wider scientific community. By sharing research datasets through established platforms [43], repositories [44], or journals [45], researchers increase the discoverability of their data, enabling others to locate and access it easily. By publishing and archiving datasets, researchers provide open access to their data, enabling others to reuse and validate their research findings. Driven by the desire of Open Culture to democratize knowledge and promote equitable access to information, Open Access promotes transparency and fosters collaboration and reproducibility in the scientific community. It allows for verifying research outcomes and supports the promotion of accessibility.

Commonly accepted formats enhance the datasets’ interoperability [46]. By adhering to established data standards, researchers enable easy integration, comparison, and combination of datasets from different sources. This is especially important in multimodal music datasets encompassing various data types. Standardized data formats facilitate data harmonization and interoperability, ensuring that data can be efficiently utilized across diverse applications.

It is important to mention that publishing and archiving multimodal music datasets has challenges rooted in the diverse data types and the potential dataset size. Often encompassing audio and video recordings, textual data, and images, these datasets pose a significant challenge due to their substantial data volume. High-resolution images, lengthy audio tracks, and extensive textual content contribute collectively to dataset sizes that strain storage capacities and retrieval speeds. Choosing an appropriate storage infrastructure is pivotal, depending on the dataset scale, access requirements, and budget constraints.

Discussion

Navigating legal frameworks raises many questions and challenges, particularly concerning music AI systems. This includes concerns about compliance with copyright laws, especially in the field of AI-generated musical content [14][10][15]. Obtaining the necessary licenses for use and distribution fosters responsible content usage and advances music research without infringing on creators’ rights. Navigating privacy complexities and protecting participants’ sensitive information requires legal awareness to mitigate violation risks.

The development and application of multimodal music datasets also underscore a range of ethical concerns, emphasizing the need for meticulous attention to responsible and transparent research practices. These include practices like prioritizing informed consent and anonymization in user-generated data collection. Furthermore, respecting artists’ information and acknowledging their work is essential. Mitigating unwanted biases in source data, and ensuring fair representation and cultural sensitivity is essential for inclusive data collections. At the same time, providing clear documentation on dataset origins and terms, and regularly assessing ethical implications guarantee data accessibility.

While this paper highlights the importance of responsible data management in the context of multimodal music datasets, it is important to acknowledge the complexity and variability of legal frameworks across different jurisdictions, which can make it challenging to create universally applicable guidelines. Additionally, the rapid pace of technological advancements means that our recommendations may need continual updates to stay relevant. Furthermore, implementing ethical and legal standards can be resource-intensive, requiring significant time and financial investment, which may not be feasible for all organizations. There is also the potential for unintended biases in datasets that even rigorous documentation and ethical oversight may not completely eliminate. These limitations highlight the need for ongoing research and adaptation in this evolving field.

Despite these challenges, our paper contributes to advancing knowledge in several ways. It provides a comprehensive overview of the current legal, ethical, and practical challenges in managing multimodal music datasets. By emphasizing the importance of obtaining proper licenses, navigating privacy concerns, and addressing ethical issues, we offer a framework for responsible data management. Additionally, we propose the development of standardized evaluation frameworks and interdisciplinary approaches to enhance the usability and responsible use of these datasets.

Future work should focus on further regulating generated music while restricting the overuse of AI systems in the creative process. It is also necessary to concentrate on assisting data managers and users, providing clear guidelines with justified recommendations for data organization. Paying attention to privacy, unwanted bias mitigation, inclusivity, and fairness may reduce the gap and inconsistencies across different legal frameworks. Developing standardized evaluation frameworks and encouraging data sharing are crucial for ensuring multimodal music datasets’ usability and responsible use. Additionally, interdisciplinary approaches, quality control mechanisms, and scalable storage infrastructure could contribute to the effective management, analysis, and needed future advancements in multimodal music datasets.

Conclusions

Managing multimodal music datasets involves navigating legal and ethical challenges, underscoring the importance of adhering to intellectual property rights and moral standards. Legal considerations encompass obtaining necessary licensing agreements and staying informed about legal developments, while ethical aspects involve recognizing creators’ moral rights and providing transparent terms of use when sharing datasets. Both legal and ethical frameworks play crucial roles in responsibly and lawfully addressing copyright complexities. Addressing these complexities responsibly and lawfully necessitates staying informed about legal developments, and providing transparent terms of use when sharing datasets.

Managing multimodal music datasets presents a heightened level of complexity due to several factors. Firstly, managing copyrights for each unique data type can be intricate, necessitating distinct approaches for different data types. Moreover, multimodal music datasets involve diverse stakeholders, including performers, photographers, researchers, and engineers, which requires establishing clarity regarding ownership and obtaining appropriate licensing agreements. Additionally, the emergence of AI-generated music data presents a special case that is not yet fully regulated, adding a layer of complexity to copyright management. Furthermore, ensuring transparency and addressing bias in multimodal music datasets is more complex due to the unique nature of each data type, making the formulation of fair guidelines specifically tailored to them a complex endeavor.

Data collection forms the foundation of any AI system, and adhering to FAIR principles at this stage is crucial. It is essential to ensure that the data collected is accurate, representative, and free from bias, reflecting a diverse range of individuals and populations to prevent discrimination, exclusion, or marginalization. This comprehensive and unbiased approach to data collection helps prevent the amplification of biases during the training and deployment stages of AI models. Furthermore, data publishing and archiving are crucial in promoting transparency, collaboration, and reproducibility. By sharing multimodal music datasets through open-access platforms and repositories, researchers enhance the discoverability of their data, enabling others to easily access, reuse, and validate their research findings. Open access to published and archived datasets fosters collaboration and supports the FAIR principle of promoting accessibility and transparency in the scientific community.

Ethics Statement and Acknowledgments

The authors declare no conflicts of interest. The work of this paper was state-funded. Further details are anonymized for submission.