Automatic tune family detection in a corpus of Irish traditional dance tunes

Tune family detection is a long-standing research area in musicology; it also aligns closely with the task of cover song detection, which is an important area of work in Music Information Retrieval (MIR). We present a large-scale experiment comparing the performance of multiple pattern-based tune similarity metrics, applied to a tune family detection task on The Session, a large corpus of Irish traditional dance tunes in symbolic notation. To allow for quantitative results analysis, ground truth tune family membership annotation is added for a test subset of 10 tune families, comprising 314 tunes in total. The corpus is preprocessed via a custom pipeline: feature sequences are extracted from each tune, from which patterns are extracted via n-grams. Three similarity methods are tested: 1. Motif : A novel, musicologically-informed approach, based on counting and weighting occurrences of similar n-gram patterns between tunes; 2. Incipit and cadence: An extended implementation of a traditional musicological incipit search; 3. TF-IDF: A pairwise Cosine similarity between n-gram pattern TF-IDF vectors for each tune in the corpus. The experiment outputs 30 distinct results sets, based on variations of these three core methodologies. The performance of each results set is quantitatively measured via Mean Average Precision; the best-performing results are selected for visualisation, analysis and discussion. Conclusions indicate that the work is a significant contribution to the computational study of Irish traditional music, satisfying the research questions, and opening up multiple lines of enquiry to be pursued through future work.

Funder

European Union Horizon 2020

Publisher

University of Galway

Rights

CC BY-NC-ND

cbn

Collections

University of Galway Theses (Research Masters Theses)