Dr. Bryan Duggan's Pages @ DIT

Bryan Duggan, "Machine Annotation of Traditional Irish Dance Music", PhD Thesis, Dublin Institute of Technology. (2009) |pdf| |abstract|

Estimates put the canon of traditional Irish dance tunes at at least seven thousand compositions. The literature attributes this to the geographic isolation of rural communities which developed their own repertoire of tunes. Musicians playing traditional music have a personal repertoire of up to one thousand tunes. Given this diversity, a common problem faced by musicians and ethnomusicologists is identifying tunes from recordings. This is evident even in the number of commercial recordings whose title is gan ainm (without name). The work presented in this PhD thesis attempts to solve this problem by developing a Content Based Music Information Retrieval (CBMIR) system adapted to the characteristics of traditional Irish music. The thesis includes a comprehensive review of the domain of traditional Irish music and presents three chapters of related work in the fields of feature extraction, melodic similarity and music information retrieval. A system is presented called MATT2 (Machine Annotation of Traditional Tunes) whose primary goal is to annotate recordings of traditional Irish dance music with useful metadata including tune names. MATT2 incorporates a number of novel algorithms for transcription of traditional music and for adapting melodic similarity measures to expressiveness in the playing of traditional music. It makes use of an onset detection function developed for the playing of traditional music on woodwind instruments such as the concert flute and tin-whistle. It uses a novel transcription algorithm based on Brendan Breathneach’s observations about the transcription of traditional Irish music which provides transposition invariance for the keys and modes used to play traditional music. It incorporates a new algorithm for dealing with ornamentation notes and accommodating "the long note" in traditional music called Ornamentation Filtering. It makes use of publicly available collections of traditional music available in the ABC notation. It uses a matching algorithm tolerant to errors which aligns short queries with longer strings from a corpus of known tunes, meaning that the algorithm can match entire tunes, incipits and phrases from any part of tune with equal success. The matching algorithm has also been adapted to take account of phrasing and reversing effects. A new algorithm is presented called TANSEY (Turn ANnotation from SEts using SimilaritY profiles) which annotates sets of tunes played segue as is the custom in traditional Irish dance music. The work presented in this thesis is validated in experiments using 130 real-world field recordings of traditional music from sessions, classes, concerts and commercial recordings. Test audio includes solo and ensemble playing on a variety of instruments recorded in real-world settings such as noisy public sessions. Results are reported using standard measures from the field of information retrieval (IR) including accuracy, error, precision and recall and the system is compared to alternative approaches for CBMIR common in the literature.
[hide]

Research

2009

2003