Assistant Professor at IMT Atlantique (BRAIn team)
PhD Graduate in Signal Processing, Computer Science Engineer
TalksI recently gave a talk about the work done in my PhD at the Centre for Digital Music (C4DM), Queen Mary University of London, untitled "Unsupervised Barwise Music Compression for Pattern Uncovering and Structural Segmentation".
Structural Segmentation of MusicMy PhD focuses on structural segmentation of music, that is, techniques to retrieve a simplified organisation of a song.
The core idea in structural segmentation is that music is based on a limited set of segments which are repeated across the song (potentially with alteration).
As presented above, a classical musical structure is the alternation of verses, chorus and solos.
Structure is multi-dimensional, as every segment can be divided in shorter segments, musical phrases or lines, and, conversely, segments may be gathered.
For this task, I have mainly studied three paths (for now):
- Nonnegative Tucker Decomposition (click to expand)
- - In short: it's a tensor factorization technique, similar in some way to NMF (Nonnegative Matrix Factorization), which can extract patterns in the song.
- Barwise Music Compression (click to expand)
- - In short: it uses linear and non-linear compression methods to compress the different bars of the song, and then infers the song structure with these compressed representations.
- Polytopic representation of music (click to expand)
- - In short: it's a paradigm which defines a local cost on a musical segment. The goal is then to minimize the sum of these local costs at the scale of the song.
Music in generalFinally, I am passionate about music (convenient for this subject !).
I also play drums (since 15 years or so), and I recently started bass (4 years ago).
Even if my musical practice is not directly related to my research, it is a huge part of my life, and, consequently, has a strong impact on my interest for the field (and on how I conceive music).