Monthly Talk: Axel Marmoret (IMT Atlantique, FR)

Abstract:

Music audio signals are rich and complex, making the extraction of meaningful content (notes, instruments, structure) a significant challenge. While deep learning dominates Music Information Retrieval (MIR), its reliance on extensive data collections, computing infrastructure, and human annotations often constitutes a major bottleneck. Furthermore, most deep learning models remain challenging to interpret and steer. Low-rank factorization methods offer a compelling alternative: they provide elegant, unsupervised tools to blindly untangle audio mixtures and uncover interpretable musical structure directly from the signal.

This presentation will explore the ongoing relevance and trajectory of low-rank factorization for music analysis, centered around two contributions: a standardized benchmarking toolbox for evaluating low- rank models on MIR tasks, and an unsupervised tensor-based method for musical pattern extraction, with potential use for interpretable and steerable music recomposition.

Together, these demonstrate that low-rank matrix and tensor factorization remain highly relevant in the deep learning era, given their computational efficiency and interpretability. In particular, such tools can prove very useful in low-resource contexts, such as traditional and historical music analysis, where data is very scarce.