Join us

🇫🇷 Français
🇬🇧 English

PhD Thesis 2020-23 – Computer music, machine learning

Musical creation: Generation guided by the learning of texture and perception

  • PhD Thesis 2020-2023
  • Location: Amiens (Laboratoire MIS, Université de Picardie Jules Verne)
  • Open for applicants.Apply preferably before end of April, and in any case before May 10
  • Remote interviews at the end of April / beginning of May, then selection by the doctoral school
  • Supervisors and contacts: Florence Levé (MdC HDR MIS, UPJV), Louis Bigo (CRIStAL)

Context

The music is made of melodies and harmonies structured in time and heights. The musical score formalizes a set of sounds and is one of the main means to transmit, exchange and preserve musical works in the West. Today, computer-based approaches combining musical expertise, algorithmic techniques and learning techniques make it possible to classify, analyze and generate music [1,2]. However, most of these approaches, traditional or using AI methods, consider the ‘note’ as the main object and then model the harmony or structure, concepts well studied in musicology as in computer music. A more abstract but very musical notion, the ‘texture’, makes it possible to describe more finely - more musically - the characteristics of a piece or song. A solo, a rhythmic beat, a homorhythmy, an accompaniment pattern, an imitation, all these textures are found in many musical styles and play a big role in musical perception.

Goals

The first studies of texture modeling are recent in computer music [3,4]. This thesis subject in the SDMA team of the MIS laboratory, in collaboration with the Algomus team of the CRIStAL laboratory, proposes to explore the musical texture in several aspects:

  • from the angle of modeling, encoding, analysis and generation, relying on algorithmical techniques as on learning on large corpora.
  • from the angle of musical perception: the choices in the textures as well as the music generated will be confronted with music perception data to serve as a basis for future applications in musical cognition or music therapy. The intuition is that taking into account the texture will allow the analyzis and the generation of music more original and more coherent, and that the different textures will be able to induce different states in the listener, which will be able to guide the choice or the generation of music in a therapeutic setting.
  • application to music therapy: music therapy can be a valuable help to improve the movements of persons living with Parkinson’s Disease [5]. The music generated will be proposed for listening as part of an evaluation of the approach for patients with Parkinson’s disease.

The thesis will be part of a network of collaborations woven by MIS and CRIStAL in recent years, whether it be academic collaborators in computer music as well as in musicology, music teachers, and R & D laboratories. The repertoires studied will be defined with the candidate and may for example relate to classical and romantic periods (string quartets by Mozart or Haydn, piano sonatas, symphonic music) or a more recent repertoire (songs, jazz transcriptions).

References

PhD Thesis 2020-23 – Computer music, text algorithmics

Indexing of melodic and harmonic patterns

  • PhD Thesis 2020-2023
  • Location: Lille (CRIStAL, CNRS, Université de Lille, Villeneuve d’Ascq), collaboration with the University of Rouen (LITIS, University of Rouen)
  • Open for applicants. Apply preferably before end of April, and in any case before May 4
  • Remote interviews at the end of April / beginning of May, then selection by the doctoral school
  • Supervisors and contacts: Richard Groult (MdC MIS, University of Picardie Jules-Verne), Mathieu Giraud (DR CNRS, CRIStAL, University of Lille, Thierry Lecroq (Pr. LITIS, University of Rouen)

Algomus is a computer music team, from the CRIStAL laboratory (UMR CNRS 9189, University of Lille) and collaborates with the MIS (UPJV, Amiens). Algomus is interested in computer analysis of musical scores. The TIBS team (LITIS, University of Rouen) works in bioinformatics, and more generally is specialized in text algorithms and indexing structures.

Context

Repeats and contrasts make music. The “musical patterns” are very present in many styles of tonal western music (baroque, classical, romantic, jazz, pop…).

A pattern can be seen as a melody (series of notes), but better models link the pattern to the underlying harmonies [Lerdhal 1988, Jansen 2013]. Today, learning-based methods allow us to learn what a musical pattern is. However, these methods do not allow efficient queries to compare a pattern with large corpora. Seed-based heuristics have already been proposed for some queries [Martin 2012]. The last twenty years have seen the emergence of numerous models in text algorithms for efficiently indexing and searching for symbolic sequences, including approximate ones, in particular by structures based on Burrows-Wheeler transform (BWT) [Adjeroh 2008].

Goals

The goal of the PhD thesis is to design, implement and test indexing structures adapted to musical patterns in symbolic scores. After a bibliography on indexing and musical patterns, the thesis may seek, for example, to adapt the BWT to search for “diatonic” patterns and for patterns described by intervals between several voices. Special care will be taken to complexity in time and in memory of the proposed solutions. The thesis will also investigate approximate searches.

The proposed algorithms will be implemented and tested on musical corpora which will have to be defined, whether in baroque / classical / romantic music or in jazz or pop music. The results will be discussed with music theorists who the team collaborates with. The goal is to publish these models and their evaluation in conferences and / or journals of computer music as well as theoretical computer science.

The PhD student will also seek to make the results usable by people analyzing music (teachers, students, composers). For this, the methods will be tested and disseminated within the Dezrann music platform developed in the Algomus team and used by music teachers and classes in the Hauts-de-France region.

Profile of the candidate: MSc in theoretical computer science, algorithms, complexity, indexing structures. Musical knowledge and practices, ideally with knowledge of harmony and / or analysis.

References