Key Distributions as Musical Fingerprints for Similarity Assessment

1
Key Distributions as Musical Fingerprints for
Similarity Assessment

• IEEE ISM 2005

Arpi Mardirossian and Elaine Chew Integrated
Media Systems Center, Epstein Department of
Industrial and Systems Engineering University of
Southern California Viterbi School of
Engineering mardiros, echew_at_usc.edu
2
Outline

• Introduction
• Method outline
• Segment
• Determine key
• Obtain key distribution
• Compare two pieces
• Experiment
• Conclusion
• Discussion

3
Introduction

• This paper present a method that may be used to
  determine whether two pieces of music are
  similar.
• Key patterns and exploits the unique combination
  of keys within a musical piece to create a
  musical fingerprint.
• Using correlation coefficient to assess
  similarity of the pieces of music.

4
Main Prograss

1. Segment
2. Determine Key
3. Obtain Key distribution
4. Compare two pieces

5
Segment

• Segment the music into m segments.
• m is a user defined variable.
• In the experiments, its concluded that the
  length of each segment should be 0.125 to 0.625
  sec. So that we can use this result to decide m.

6
Determine Key

• For key recognition, this paper use the Spiral
  Array Center of Effect Generator Algorithm.
• The Spiral Array is a mathematical model for
  tonality that represents tonal elements, such as
  pitch classes and keys, using a set of nested
  helixes.

7
Determine Key

• The collection of pitches in each segment is
  mapped to their corresponding positions in the
  pitch class spiral using a pitch spelling
  algorithm
• The aggregate position of these positions is
  obtained by weighting each pitch class
  representation by its proportional duration in
  the segment.
• The key is then determined through a nearest
  neighbor search for the nearest key presentation
  on the major and minor key helixes.

8
Determine Key

• D duration
• P pitch position
• Compute the dist. from ci to each key.

9
Determine Key
10
Obtain key distribution

• We refer to the set of keys of the segments as
  the m dimensional vector K k1, k2, , km.
• Let P be the set of all possible keys.
• We next create a key frequency vector
• F f1, f2, , f55
• where fi represents the number of times an
  element of K is equal to the ith element of P.

11
Compare two pieces

• compare the F vectors of two musical pieces.
• r -1(representing patterns that are complete
  opposites)
• r 1(representing patterns that are exactly the
  same).
• x is the values contained in F of the first piece
  
• is the average of x.
• y is the values contained in F of the second
  piece.
• is the average of y.

12
Experiments
R 0.094
R 0.909
13
Experiments
Compare the distributions of correlation
coefficients for comparisons of pieces from the
same set and from different sets. Result is
normalized gt sum 1.
14
Conclusion

• From the result of experiment, key distribution
  can be used to develop musical fingerprint for
  similarity assessment.

15
Discussion

• Music is sequential and contextual in nature, but
  this paper ignores sequence information and that
  each segment is treated in isolation of its
  neighbor.
• If it is possible that they also ignore the key
  transition and the similar pieces are detected as
  dissimilar?