Index of /~echew/papers/ISMIR7

 Name                    Last modified      Size  Description
 Parent Directory                             -   
 mc-ISMIR0613_Paper.pdf  03-Nov-2006 10:44  490K  
 reference.txt           03-Nov-2006 10:46  452   

In this directory are the PDF files for the paper titled

"Music Summarization Via Key Distributions: Analyses of Similarity Assessment Across Variations"
by Arpi Mardirossian (mardiros@usc.edu) and Elaine Chew (echew@usc.edu)

The results were presented at the 
Seventh International Conference on Music Information Retrieval
Victoria, B.C., Canada.  October 8-12, 2006.

The paper is published in the
Proceedings of the 7th ISMIR Conference
Click on reference.txt for the BibTeX reference.

The conference website is at 
http://ismir2006.ismir.net

THE COMPLETE PAPER, text with figures, can be viewed as a PDF document.
Click on mc-ISMIR0613_Paper.pdf if you wish to view the paper in PDF format.

--'--,--'--,--'--,--'--,--'--,--'--,--'--,--'--,--'--,--'--,--'--,--'--

"Music Summarization Via Key Distributions: Analyses of Similarity Assessment Across Variations"
by Arpi Mardirossian (mardiros@usc.edu) and Elaine Chew (echew@usc.edu)

ABSTRACT: This paper presents a computationally efficient method for
quantifying the degree of tonal similarity between two pieces of
music. The properties we examine are key frequencies and average time
in key, and we propose two metrics, based on the L1 and L2 norms, for
quantifying similarity using these descriptors. The methods are
applied to 711 classical themes and variations over 71 variation sets
by 10 composers of different genres. Quantile-quantile plots and the
Kolmogorov-Smirnov measure show that the proposed metrics exhibit
strongly distinct behaviour when assessing pieces from the same
variation set, and those that are not. Comparisons across variation
sets by the same composer, and comparisons of pieces by different
composers although result in similar distributions, are derived from
fundamentally different underlying distributions, according to the K-S
measure. We present probabilistic analyses of the two methods based on
the distributions derived empirically. When the discrimination
threshold is set at 55, the probabilities of Type I and Type II errors
are 18.41% and 20.56% respectively for Method 1, and 15.72% and 22.94%
respectively for Method 2.  Method 1 has a success rate of 99.48% when
labeling pieces as dissimilar (not from the same variation set), while
the corresponding rate for Method 2 is 99.45%.

Keywords: Music similarity, similarity assessment, music
representation, music summarization, key distribution, pitch, music
information retrieval.