Title: Exploration of Music Collections with Audio Embeddings

Supervisors: Dr. Xavier Serra and Dr. Dmitry Bogdanov

Jury: Dr. Fabien Gouyon (Pandora), Dr. Frederic Font (UPF), Dr. Eva Zangerle (University of Innsbruck)

Abstract:

Music recommendation systems (RecSys) are integral to modern music streaming services. While there is much research on many aspects of RecSys, there is not enough research on exploration and discovery that contributes to long-term user retention. After conducting an anonymous survey, we identify that the exploration and rediscovery of the personal collections in particular needs improvement.
To address this, we take advantage of music tags (genre, moods) and use deep auto-tagging systems to construct latent spaces. We investigate different architectures, datasets, layers, and projections and how they affect the perceived similarity of nearest neighbors. Finally, we present a novel web interface to visualize music collections using audio embeddings.
We evaluate the proposed solution via semi-structured user interviews and conclude that it provides an excellent alternative to existing solutions. We believe that the contributions of this work enable more research and industry solutions for music exploration and discovery.

This thesis defense will take place in hybrid format. To attend online use this link (meeting ID 921 5847 2306). The microphone and camera must be turned off, and the online access will be unavailable after 30 minutes from the start of the defense.

Video: https://youtu.be/-b_CsR_WggM