Morgan Buisson

prof_pic.jpg

Hello! I’m a postdoctoral researcher at CNRS in Nantes, working with Dr. Vincent Lostanlen on Multi-Resolution Neural Networks (MuReNN) for audio. I completed my PhD at the ADASP Group at Télécom Paris, where I was jointly supervised by Prof. Slim Essid and Brian McFee from New York University.

My research lies at the intersection of machine learning, signal processing, and audio data analysis, with a particular focus on Music Information Retrieval (MIR).

Academic Journey

  • 2014–2019: I earned a master’s degree in Applied Mathematics from the Institut National des Sciences Appliquées de Rouen, France.
  • 2018–2019: I was a graduate exchange student at École Polytechnique de Montréal, Canada.
  • 2020–2021: I completed the master’s program in Sound & Music Computing at the Music Technology Group (MTG) at Pompeu Fabra University in Barcelona, Spain.
  • 2021–2024: I was a doctoral researcher at the Audio Data Analysis and Signal Processing Group at Télécom Paris (Institut Polytechnique de Paris). My doctoral research focused on Music Structure Analysis. You can access my thesis manuscript here.
  • Summer 2025: I completed a research internship at Spotify, supervised by Dr. Rachel Bittner, where I worked on music summarization.

Doctoral Research Focus

Estimating song structures is a crucial task in Music Information Retrieval, yet it presents challenges due to the ambiguity of structure annotations and the scarcity of labeled data. My PhD research addressed these issues through three main approaches:

  1. Self-supervised Learning for Music Segmentation — I developed methods that leverage prior musical knowledge, such as the hierarchical nature of music structure, to learn audio representations that enhance segmentation performance without labeled data.

  2. Graph-based Music Structure Analysis — By framing music structure analysis as a link prediction task, I applied deep graph learning techniques to achieve state-of-the-art segmentation results with minimal supervision, improving both performance and interpretability.

  3. Multimodal Learning with Language Models — I explored the connection between text and audio, utilizing language models to tackle ambiguities in music structure annotations.

Research Interests

Self-supervised learning for sound understanding: I plan to continue investigating how pseudo-tasks can help models learn rich audio representations without human annotations. This approach holds promise for scaling up learning in domains where labeled data is scarce or ambiguous.

Structure-aware generative modeling of music: Building on my PhD experience, I would like to study how discriminative models of musical structure can inform generative systems. The goal is to create music generation models that better capture how local and global elements come together to form coherent musical pieces.

Ethical AI and artist protection: A growing amount of research and financial investment is directed toward music generation technologies, often with limited consideration for the rights of artists whose work underpins these systems. I am interested in exploring more equitable approaches to AI in music.