FUNDAMENTAL RESEARCH

AI Methodologies for Music Creation and Production

One of the research areas of the Chair is the use of generative models, especially those based on deep learning, for music creation and production. Notable work has been done on transformer architectures applied to melody and rhythm generation, as exemplified by the research line of the Groove Transformer model, validated in performative environments with musician Refree. Furthermore, the use of diffusion models and style transfer techniques for audio synthesis has been investigated, leading to contributions presented at international conferences such as ISMIR and NeurIPS Music Workshop.

An example of the integration of these models into accessible tools for musicians, designed from a collaborative and user-centered perspective is "El Bongosero" interactive installation at the CCCB, a citizen science experience that allowed for the collection of valuable data to train adaptive percussion models. This work, along with collaborations with artists and festivals, has explored the potential of AI as a co-creator in real musical processes, while also generating academic knowledge and critical reflection on computational creativity.

AI Methodologies for Music Distribution

The research line dedicated to music distribution is exploring recommendation models, content analysis, and personalization through machine learning systems. Collaborations with companies like BMAT, Moises, and Deezer alow us working with real-world, large-scale data, evaluating multimodal models that combine audio, text, and metadata signals to improve information retrieval and transparency in recommendation systems.

Scientific results include studies on algorithmic transparency and responsible evaluation of recommendation systems, aligned with the requirements of the European AI Act.

Work has also been carried out on methodologies for detecting and marking AI-generated content, as part of an emerging line on traceability and regulation. Several studies conducted in this area have contributed to technical reports for the European Commission, consolidating the chair as a relevant agent in designing technological solutions compatible with emerging legal frameworks. These developments are key to ensuring that AI applied to music distribution respects creators' rights and facilitates a more equitable relationship between users, platforms, and artists.

AI Methodologies for Music Education

Research in this area focuses on developing AI models that support music learning and teaching processes, in both formal and informal settings. Techniques for automatic score recognition, performance evaluation, and instrumental study assistance have been explored. Notable examples include collaborations with institutions such as Trinity College London, MOTMO, and ESMUC, and the development of prototypes like Score Entropy Explorer and feedback systems for piano or singing students, integrating audio-symbolic analysis.

An especially relevant contribution has been the work on interactive interfaces that allow students to explore musical repertoire using natural language, combining NLP and symbolic representation. Additionally, efforts have been made in accessibility and personalized learning, promoting technologies that adapt to the user's level and needs. These methodologies are being evaluated in educational contexts and case studies on Baroque opera or classical quartets, offering a clear path to transfer scientific knowledge into practical tools with social impact.

APPLIED RESEARCH

The applied research activities of the chair are oriented toward developing concrete technological tools that address the needs of the music sector in the field of artificial intelligence. Significant progress has been made in both software development and the creation of datasets and interactive prototypes. These tools not only support fundamental research but also enable its direct transfer to professional and educational environments, reinforcing the impact of our investigations.

Software:

Essentia platform: in the context of the chair, new versions have been released incorporating functionalities to work with deep learning models, as well as tools for segmentation, rhythmic and harmonic feature extraction, and musical structure detection.

MTG-Toolbox: a collection of modular and accessible utilities for musicians, educators, and developers, aimed at facilitating work with AI models applied to tasks such as music generation, transformation, and analysis.

Data:

Freesound, the collaborative platform for openly licensed sounds, which continues to expand its user base and also incorporates AI-assisted tagging and search tools.

Over 10 specific datasets have been published in the MTG's Zenodo community. These sets include audio data, symbolic annotations, textual descriptions, and multimodal data designed for tasks such as source separation, structure analysis, style recognition, or supervised machine learning in music.

Prototypes:

MTG Demos site: interactive, online-accessible prototypes, allowing exploration of technologies such as groove generation, stylistic adaptation of melodies, sound similarity search, or automatic audio description.