news

QMUL, School of Electronic Engineering and Computer Science

Centre for Digital Music Seminar Series

Seminar by:
Axel Marmoret (IRISA, Rennes, France)

Date/time: Wednesday 16th of March, 3:00-4:00 pm

Location: Graduate Centre - Room 104, Mile End Campus & Online

Open to students, staff, alumni, public; all welcome. Admission is FREE, no pre-booking required.

Title: Unsupervised Barwise Music Compression for Pattern Uncovering and Structural Segmentation

Bio: Axel Marmoret, 25 years old, is currently doing a PhD in IRISA, Rennes (France), on the subject of Audio Structural Segmentation.

Abstract: The first part of this talk will be dedicated to the presentation of Nonnegative Tucker Decomposition (NTD), a tensor factorization technique. Applied to barwise representation of a song, NTD is able to uncover barwise audio patterns, in an unsupervised way [1, 2, 3]. These patterns can in turn be used to compose new songs [3], or as a tool for Music Information Retrieval (MIR), as in the context of Structural Segmentation [1,2]. Secondly, the talk will present linear (PCA) and nonlinear (AutoEncoders) compression techniques, applied to barwise representation of music [4]. These compressed representations are studied on the task of Structural Segmentation, where they improve the former state-of-the-art for unsupervised methods, and can achieve the level of performance of supervised state-of-the-art. Finally, this talk will present recent developments about a new compression paradigm, mixing NTD and AutoEncoders, which could be a new lead for Explainable Deep Learning in MIR. All of this work is being developed in the context of my PhD.

Presentation: [PDF Slides] [PPT Slides] [Video Recording]

[1] Marmoret, A., Cohen, J., Bertin, N., & Bimbot, F. (2020, October). Uncovering Audio Patterns in Music with Nonnegative Tucker Decomposition for Structural Segmentation. In ISMIR 2020-21st International Society for Music Information Retrieval.

[2] Marmoret, A., Voorwinden, F., Leplat, V., Cohen, J. E., & Bimbot, F. (2021). Nonnegative Tucker Decomposition with Beta-divergence for Music Structure Analysis of audio signals. arXiv preprint arXiv:2110.14434.

[3] Smith, J. B., & Goto, M. (2018, April). Nonnegative tensor factorization for source separation of loops in audio. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 171-175). IEEE.

[4] Marmoret, A., Cohen, J. E., & Bimbot, F. (2022). Barwise Compression Schemes for Audio-Based Music Structure Analysis. arXiv preprint arXiv:2202.04981.