Music similarity models applied to cover song identification and classification

• Ano de defesa: 2021
• Autor(a) principal: Bodo, Roberto Piassi Passos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Classificação de covers; Computação musical; Computer music; Cover song classification; Cover song identification; Identificação de covers; Music information retrieval; Music similarity; Recuperação de informação musical; Similaridade musical
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-27092021-104421/

The spread of digital music allowed the appearance of datasets with millions of music files. The processing of this huge number of audio files is carried out with techniques of Music Information Retrieval (MIR) that work directly with the audio content. The MIR task of most interest in this project is the modelling of Music Similarity. Our proposed approach follows this pipeline: extract audio features, aggregate local features into global features, and compute the similarities of every pair of songs from the dataset being processed. According to this approach, a triple {extractor_i, aggregator_j, distance_k} defines a music similarity model, and our main goal is to investigate the ability of similarity models to distinguish audio files from different classes. The music similarity models are also used to address specific problems such as Cover Song Identification (CSI), which is an MIR application related to Music Similarity, and the closely-related Cover Song Classification (CSC) problem. MIR-related techniques, such as Dataset Modifications and Matrix Fusion, are explored in the context of improving the results of music similarity models. This work presents several contributions, among which a comprehensive benchmark of music similarity models; the definition of new similarity matrices within CSC as a solution approach to the CSI problem; the exploration of different types of dataset modifications and an investigation of their effect on music similarity metrics; and the fusion of similarity matrices computed from secondary datasets obtained via source separation. The experiments presented produced encouraging results, indicating that the methods proposed in this thesis point towards novel approaches to Music Similarity that are worth further investigation and development.