Optimizing Measures Of Melodic Similarity For The Exploration Of A Large Folk Song Database

This investigation aims at finding an optimal way of measuring the similarity of melodies. The applicability for an automated analysis and classification was tested on a folk song collection from Luxembourg that had been thoroughly analysed by an expert ethnomusicologist. Firstly a systematization of the currently available approaches to similarity measurements of melodies was done. About 50 similarity measures were implemented which differ in the way of transforming musical data and in the computational algorithms. Three listener experiments were conducted to compare the performance of the different measures to human experts’ ratings. Then an optimized model was obtained by using linear regression, which combines the output of several measures representing different musical dimensions. The performance of this optimized measure was compared with the classification work of a human ethnomusicologist on a collection of 577 Luxembourg folksongs.

• Type: Conference
• Year: 2004
• Where: ISMIR
• Authors: Daniel Müllensiefen, Klaus Frieler
• Keywords: Folk Song Collection, ISMIR 2004, Similarity Measurements, Similarity Measures