Projet de recherche doctoral numero :8059


Date depot: 3 mars 2021
Titre: Structured Models for Written Music Processing
Directeur de thèse: Florent JACQUEMARD (Inria-Paris (ED-130))
Directeur de thèse: Philippe RIGAUX (CEDRIC (Ancien membre EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini

Resumé: Music Information Retrieval (MIR) is a multidisciplinary field concerned with the processing, organization, access and analysis of musical content in various formats such as audio recordings, symbolic performance recordings (MIDI), musical scores... Approaches developed in this area rely on different acoustic models and language models. The latter are often based on sequential (1D) or geometric (2D) representations of musical events (notes with attributes of pitch, start time, and duration). Common Western Music Notation (CWN) is a graphical format used for centuries as a crucial vector for knowledge transmission in musical practice. Although based on a relatively small number of symbols, this format is much more structured and conveys more information than the aforementioned representations. It indeed describes local and non-local relationships and a hierarchical organization of melodic and harmonic content in rhythmic groups, sentences, etc. Such information is useful to musicians for the understanding and interpretation of pieces, and can also be exploited in MIR tasks. The objective of this PhD is to study (i) structured music representations sharing fundamental properties with CWN, (ii) language models & formalisms based on such representations and (iii) their application to several MIR tasks. We shall in particular focus on the two following problems for these models: - the construction of index for fast retrieval in digital music score databases, - the definition and efficient computation of similarity metrics like edit distances. Applications considered the PhD work may include (without being necessarily limited to): The development of tools useful in the context of the analysis or edition of music scores e.g. for the comparison of score files or the quantitative evaluation of measures of notational quality or complexity; Information retrieval in databases of digital music scores, e.g. for melodic search, pattern extraction, identification of similar fragments etc; The use of our language models as an intermediate representation (and referential) in order to leverage processing tasks to heterogeneous content, in various encoding formats (MusicXML, MEI, MNX, Lilypond, Guido, kern**...). In order to motivate these applications, particular attention will be dedicated to the development of a relevant use case in computational musicology, in collaboration with musicologists at Sorbonne University. The development of collections and tools will be integrated into the base Neuma to demonstrate the practical impact and to disseminate the project's results.

Doctorant.e: Rodriguez-De La Nava Lydia