Melodic Similarity in Traditional French-Canadian Instrumental Dance Tunes

Commercial recordings of French-Canadian instrumental dance tunes represent a varied and complex corpus of study. This was a primarily aural tradition, transmitted from performer to performer with few notated sources until the late 20th century. Practitioners routinely combined tune segments to create new tunes and personalized settings of existing tunes. This has resulted in a corpus that exhibits an extreme amount of variation, even among tunes with the same name. In addition, the same tune or tune segment may appear under several different names. Previous attempts at building systems for automated retrieval and ranking of instrumental dance tunes perform well for near-exact matching of tunes, but do not work as well in retrieving and ranking, in order of most to least similar, variants of a tune; especially those with variations as extreme as this particular corpus. In this paper we will describe a new approach capable of ranked retrieval of variant tunes, and demonstrate its effectiveness on a transcribed corpus of incipits.