Music Recommendation System Based On Emotions

Abstract

Most music recommendation systems focus solely on music similarity techniques to suggest songs, disregarding the listener's emotional context. This thesis describes a system that proposes a recommendation strategy that orders songs by music annotation proximity on the Valence-Arousal emotion model. It also considers a second strategy that applies psychoacoustic models. The technique relies on a modeling via Mel Frequency Cepstral Coefficients extraction from songs. Additionally, this thesis proposes an experiment that consists on the evaluation of the quality of the recommendations according to the listener's emotional perception in music. Both recommendation strategies are then compared to understand which performs best. The results demonstrate that recommendations that use emotional annotations have a better performance, wherein the user profile and distribution of emotions selected by listeners on the bidimensional plane of emotions does not influence the evaluation of recommendations. To conclude, several suggestions on how to continue and improve this work are made, aiming the future application of this approach in an academic context and/or in a commercial product.