Deep audio representation learning for music using weak supervision

Music audio tagging is the Music Information Retrieval task of assigning one or multiple labels to an audio signal. Current state-of-the-art music taggers rely on deep learning approaches, which offer high performance but introduce challenges due to their large data requirements and tendency to over...

Descripción completa

Detalles Bibliográficos
Autor: Alonso Jiménez, Pablo
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2024
País:España
Institución:CBUC, CESCA
Repositorio:TDR. Tesis Doctorales en Red
OAI Identifier:oai:www.tdx.cat:10803/692389
Acceso en línea:http://hdl.handle.net/10803/692389
Access Level:acceso abierto
Palabra clave:Representation learning
Music Information Retrieval
Music tagging
Music classification
Deep learning
Audio processing
Aprenentatge de representacions
Recuperació d'Informació Musical
Etiquetatge d'àudio musical
Classificació musical
Aprenentatge profund
Processament d'àudio
62
Descripción
Sumario:Music audio tagging is the Music Information Retrieval task of assigning one or multiple labels to an audio signal. Current state-of-the-art music taggers rely on deep learning approaches, which offer high performance but introduce challenges due to their large data requirements and tendency to overfit. In this thesis, we propose addressing music tagging from the perspective of representation learning, which consists of designing pre-training objectives that make the learned representations suitable for several downstream tasks. In our work we investigate using representations learned by competitive music and audio tagging systems, the effectiveness of training representation models on music metadata (such as artist names and playlists) as a source of supervision, the usage of the transformer architecture for representation learning, and the adaptation of audio interpretability strategies to operate with pre-trained representations. Most of the models developed in this thesis were incorporated into Essentia,1 an open-source sound and music analysis library.