A comparison of approaches for processing the time dimension in automatic cover song identification

This master’s thesis compares different approaches for processing the time dimension in automatic Cover Song Identification (CSI). Six Time Handling Approaches (THAs) are conceptualized based on existing literature. To test the THA’s impact on cover song prediction performance, a two-phased research strategy is chosen. During the exploratory phase, song pairs are compared qualitatively, while the subsequent retrieval phase evaluates different THAs in a testing environment using various evaluation metrics. While the exploratory phase yields promising results for some THAs, the baseline employed during the retrieval phase cannot be outperformed. The results qualitatively show that longer chunk lengths in THAs typically lead to better prediction performance and are more suited for pairwise comparison than for large-scale retrieval tasks.

Grantor dc:publisher

Humboldt-Universität zu Berlin

Year dc:date.issued

2025

Author dc:creator

• Heigenhauser, Jonas

• machine learning
• information retrieval
• music information retrieval
• cover song identification
• audio analysis
• music representation
• spectrogram

Statement dc:rights

• (CC BY-NC-ND 4.0) Attribution-NonCommercial-NoDerivatives 4.0 International

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-nc-nd/4.0/

Language dc:language.iso

eng

DOI

https://doi.org/10.18452/35505

OAI record oai:identifier

https://edoc.hu-berlin.de/server/oai/request?verb=GetRecord&metadataPrefix=dim&identifier=oai%3Aedoc.hu-berlin.de%3A18452%2F36156