Get started
Work

Assister la pratique de la guitare et la composition de tablatures en musiques actuelles occidentales par apprentissage automatique

Year: 2025
Type: dissertation
Abstract: Guitarists who play Western Popular Music (WPM) are usually little trained in music theory, a great part of their learning being autonomous and informal.To learn new songs or share them to other guitarists, they will commonly resort to audio recordings and tablatures.When it comes to composing new songs, it is usually conducted directly with the guitar through jamming and experimenting, alone or in a band.This thesis aims at exploring new computational methods that could assist guitarists both in the learning and compositional phases.Learning assistance is motivated by the large quantity of guitar resources available online, and aimsat helping guitarists navigate them to ease informal learning.A first algorithmic approach was proposed to make song recommendations to learners that are conditioned bydifficulty ratings of the songs and the learner's level estimate.A second approach proposes musical features to automatically estimate the difficulty of bass and guitartablatures on multiple criteria, while guaranteeing explainability and interpretability.Both approaches are based on new datasets that are or will be shared openly to the research community.Normally, tablatures of composed songs are only written in a later stage to transcribe the songs, for memorisation and communication purposes.Other contributions of this thesis propose ways to assist with composition, the objective being to motivateguitarists to use tablatures during the composition process by augmenting tablature notation software with Machine Learning (ML) functionalities.A first model permits suggesting where to add emph{bends} in a tablature, in order to make it more idiomaticof WPM guitar playing. The ML model was integrated in an online tablature notation software to illustrate how it could fit guitarist-composers' workflow.Other contributions focus on guitar accompaniment parts.A model for suggesting a guitar chord position, given the previous one, is introduced and tested with aset of newly introduced quantitative metrics.This work was used as a first step that led to the development of a transformer and a rule-based model that generate possible continuations of an existing bar of rhythm guitar, in tablature format. Both models were evaluated quantitatively, but also qualitatively by external participants through an online survey.Finally, another transformer model is proposed for generating bass tablatures, given an existing rhythm guitar track, to assist guitarist-composers who are not bass players.
Language: English
Open Access status: green
OpenAlex ID: W7125087383
Cites:
Cited by:
Related to: