Skip to Main content Skip to Navigation

Jusqu'où les goûts musicaux sont-ils prédictibles par l'intelligence artificielle ?

Nicolas Dauban 1 
Abstract : This thesis focuses on the use of artificial intelligence (AI) for music recommendation as a prediction of music tastes. Our research focused on the study of musical tastes, the process of music recommendation and the place of humans in this process. We were interested in the relevance of the data used for music recommendation: behavioral data on the one hand, often supposed to reflect user affinities; song descriptors, used for content-based recommendation on the other hand. The proposed approach was based in particular on work in the human sciences and musicology. The issues are focused on the prediction of musical tastes, in a field where there are many uncertainties around the explainability of these tastes: - What elements influence musical tastes? What are the links between musical genres and listeners' tastes in music? - Is the music recommendation a taste prediction? How is behavioral data interpreted by streaming platforms? Does the behavior really reflect musical tastes? - Are the acoustic parameters learned by a convolutional neural network for automatic genre recognition relevant for the prediction of musical tastes? What determining factors in musical affinity is a convolutional neural network able to identify in a spectrogram? The first chapter presents our study of musical tastes, their links with genres and the contextual parameters that can influence listening to music. This study links conclusions from our bibliography and data from a streaming platform. In the second chapter, we looked at the data, methods, and evaluation modes used by streaming platforms for recommendation. We have also proposed a method to link acoustic parameters to a human categorization, for a personalized music recommendation. The third chapter presents a taste prediction method based on the conclusions of the previous chapters. This method is based on pre-trained deep convolutional neural networks for genre prediction. An offline evaluation process brought satisfactory results (more than 70% accuracy) which were confirmed by a human evaluation: 100% success on a third of the volunteers, more than 77% average accuracy. Finally, a qualitative study of the results provides ideas for improvement of the proposed system.
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Tuesday, December 7, 2021 - 4:52:35 PM
Last modification on : Monday, July 4, 2022 - 9:11:55 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03469458, version 1


Nicolas Dauban. Jusqu'où les goûts musicaux sont-ils prédictibles par l'intelligence artificielle ?. Intelligence artificielle [cs.AI]. Université Paul Sabatier - Toulouse III, 2021. Français. ⟨NNT : 2021TOU30082⟩. ⟨tel-03469458⟩



Record views


Files downloads