Compacting Music Signatures for Efficient Music Retrieval

1
Compacting Music Signatures for Efficient Music
Retrieval

• Bin Cui, H. V. Jagadish, Beng Chin Ooi and
  kian-Lee Tan
• Conference of Extending Database Technology, 2008

2
Outline

• Introduction
• The design of MUSIG
• Performance
• Conclusion
• My thought

3
Introduction

• Using compact music signatures for efficient
  content-based retrieval
• The music file can be represented by a sequence
  of notes
• Each music file is split into a set of segments
• Similar segments are clustered together
• Find database sequences of the query is a
  subsequence
• Content-based retrieval
• Acoustic and symbolic retrieval

4

• Symbol retrieval
• Represent music as a sequence of symbols
• Each symbol represent a musical sound
• It is hard which retrieval by content requires
  sub-string matching

5

• q-gram
• Each string is divided into segments of length q
• It seek database objects that include all of the
  q-grams in the query
• Approximate matching is difficult

6
The design MUSIG

• Segment the Music
• The length of the window is 6.
• The window sliding step size is 2.

7

• Each beat as one dimension and the pitch value as
  key of the corresponding dimension
• Each segment can be represented by a
  multi-dimensional vector
• For example
• Music piece (8, 10, 8, 13, 12, 12, 8, 10, 8, 15,
  13, 13)
• 4 segments (8, 10, 8, 13, 12, 12), (8, 13, 12,
  12, 8, 10), (12, 12, 8, 10, 8, 15), (8, 10, 8,
  15, 13, 13)

8
The Case for Signature Method

• Drawback
• It incurs a high computational cost
• It is difficult to control the size of each query
• Small window size more candidates
• Large window size degrade the query performance
• As a music query may start anywhere in a music
  melody
• Use small window sliding step size
• Higher storage overhead and processing cost

9
Generate the music signature

• To generate music signatures
• Split a music file into a set of segments using a
  sliding window
• Cluster all the segments obtained from the music
  database
• Address substring approximate match
• The music signature
• Each music file can be represented by a single
  high-dimensional vector

10

• Using cluster method
• The number of dimensions is the number of
  clusters
• It employ the K-means clustering scheme
• For example
• Music signature S1 is (1, 2, 5, 0, 2)

11
Query Algorithm

• There are a couple of issues to consider
• potential phase matching problem and
  approximation
• Match Score
• Using the score function to rank the returned
  results

12
Matching algorithm
13
Performance

• Parameter
• Window size (W) 4, 8, 16, 32
• Window sliding step size (WS) 1, 2, 3, 4
• Signature dimensionality (D) 100, 200, 300, 400,
  500
• Length of music query (L)1/16, 1/8, ¼, ½
• Answer rank 1, 5, 10, 20

14
Effect of window size and window sliding step size
15
(No Transcript)
16
Conclusion

• To generate the music signature
• Adopt a clustering approach
• It designed a scoring function to determine the
  match score between a music query and a music
  data
• The music signature is capable of distinguishing
  differences between music
• The number of dimensions is high

17
My thought

• Repeating pattern and clustering method
• Using tree structure