Audio Representation Learning of Musical Works

Sammanfattning

Version identification is a problem in Music Information Retrieval, where the goal is to link differentaudio recordings to the same underlying musical work. This is relevant for large music catalogs,where metadata can be incomplete, inconsistent, or incorrect. A song may appear as a studiorecording, live version, acoustic cover, or radio edit, causing recordings to differ in tempo, timbre,arrangement, and instrumentation despite belonging to the same work. This thesis investigates whether content-based audio representations can improve version iden-tification compared with a traditional chroma-based method. Three methods are compared:Chroma Energy Normalized Statistics (CENS), PANNs CNN14, and the Audio SpectrogramTransformer (AST). The two neural models were adapted to metric learning, where audio record-ings are converted into vectors and compared based on similarity. The results show that AST performs best among the evaluated models, achieving a Top-1 re-trieval accuracy of 45.8%, compared with 30.3% for CNN14 and 27.7% for CENS. This suggeststhat machine learning-based, content-based audio representations are a promising direction forversion identification and may become an important complement to metadata in large music cat-alogs. However, challenges remain, including incorrect top-ranked matches, highly varied works,and computational costs when applying these methods at scale.