Méthodes d'Apprentissage Profond pour l'Analyse de Structures Musicales : Adresser le Manque et l'Ambiguïté des Données

Audio-based music structure analysis involves automatically identifying the different sections of an audio recording. Because various factors can influence human listeners’ perception of a music piece’s inner organization, musical structure is, by nature, subject to some ambiguity and subjectivity.In this dissertation, we focus on these characteristics of musical structure. We explore different research directions that explicitly incorporate them in the analysis process. The methods proposed in this thesis leverage different levels of supervision, either from a simple heuristic linked to the implicit timeline of music, from information contained in the input signal, or from data provided with structural annotations.First, we focus on the preliminary stage of music structure analysis, extracting audio features from the input signal. We opt for self-supervised representation learning approaches that can leverage large quantities of unlabelled music audio data. On the one hand, we propose a method directly inspired by the hierarchical aspect of music structure. On the other hand, we investigate another self-supervised strategy that uses the notion of repetition as an inductive bias to inform the frame sampling process.We then consider the scenario where we are given a set of music tracks for which human listeners have annotated the structure. In such a setting, we investigate how pre-trained audio representations may be adapted to such limited data corpora and highlight cases where the use of annotations can hurt segmentation performance.We finally introduce a fully-supervised approach to music structure analysis based on graph neural networks. The core of the method lies in extracting link features between time instants within the input track, which provide useful structural cues to predict segment boundaries and section labels.Our methods are evaluated on standard datasets for music structure analysis and systematically compared with recent state-of-the-art systems. Overall, the results suggest that explicitly targeting the issue of bias in structural annotations can help design more robust music structure analysis systems, which perform competitively against previous work even in limited training data situations.