Agglomerative Clustering for Audio Classification using Low-level Descriptors
Résumé
Mainly inspired by the work of Geoffroy Peeters, Diemo Schwarz and Grégoire Carpentier in the field of music information retrieval (MIR), this document presents the work that I have achieved in the frame of Ircam Cursus 2 (Specialized Training in Composition, Research and Music Technology) under the guidance of Mikhail Malt and Hèctor Parra in 2015-2016. The main lead of this project was to elaborate a computerized classifier for large corpuses of sounds. In other words, the idea was to be able to organise and re-organise easily a database under specific constraints or concepts that could be useful in preparation of scoring a musical piece. Although parts of this idea have been investigated to achieve different tasks such as corpus-based concatenative synthesis [Schwarz, 2006], musical genre recognition [Peeters, 2007] and computer-assisted orchestration [Carpentier, 2008], the challenge remained, and still remains, to find a way to adapt this approach for compositional purposes; not only to generate material but mostly to analyse, to explore and to understand the full potential, inside out, of a sound material. That may be seen as a kind of audio data mining applied to computer-aided composition. As the title of this document reveals some pieces of answer to this interrogation, the following explanations aim at unfolding the algorithmic structure in order to examine the methodology, to discuss a few important co-lateral problematics and to analyse different clustering results. The objective is also to take a look at the influence of this research on a first artistic outcome: Alors que le monde est décomposé, for piano & electronics - performed by Wilhem Latchoumia, to expose the limitations of this approach in relation to specific artistic goals and finally, to discuss ideas for future development.
Origine : Fichiers produits par l'(les) auteur(s)