A Music Retrieval Method based on Distribution of Feature Segments

1
A Music Retrieval Method based on Distribution of
Feature Segments

• Kazuhisa Ono, Yu Suzuki, Kyoji Kawagoe
• IEEE International Symposium on Multimedia 2008

2
Outline

• Introduction
• Music Retrieval based on Distribution of Feature
  Segments
• Extract Feature Segments
• Calculating Similarity
• Experimental Evaluation
• Conclusion
• My thought

3
Introduction

• In common music retrieval methods
• A problem is that several features in the music
  are ignored.
• It calculates the importance of each feature in
  the music.
• Compare the importance of features between the
  query and the retrieval target
• It is quite similar to the idea in textual
  information retrieval techniques.

4
Music Retrieval based on Distribution of Feature
Segments
5
Extract Feature Segments

• If define a fixed length for each feature segment
• It will ignore several phrases
• Because the length of each feature segment in the
  music is not the same.
• Polyphonic data is time series data.
• Extracting flexible feature segments from the
  music is difficult.
• To solve this problem
• The sequential time series data in the music
  correspond to string in the text.

6

• It extract segments that have similar feature
  values for the songs

7

• Acoustic features
• Frequency distribution and average amplitude as
  feature values for the music
• Frequency distribution
• Difference in timbral texture
• Average amplitude
• Difference in impressions based on the strength
  of the sound

8
Feature Segments by Connected Frames

• It divide music into fixed frames, and connect
  the frames that are similar between one piece of
  music and other music pieces
• Define the connected frames as similar segments
• Set 25 msec as the length of the frame and 10
  msec as length of the frame shift

9
Classification for Feature Segments

• The average amplitude and the frequency
  distribution are values in different kinds of
  music
• Reduce the number of different kinds of feature
  segments
• Compare these feature segments with the feature
  segments that have the same length L.
• Merge the feature segments. (why?)

10
Calculating Similarity

• It utilize a TF-IDF algorithm for the vector
  retrieval model in text retrieval in music
  retrieval method.
• tkl is the frequency of kth feature segment wk in
  lth music Sl
• The frequency of wk in the query as weights qk
• Generate feature vectors that consist of each
  weight as elements

11
Experimental
12
Conclusion

• It presented a music retrieval method
• Based on the distribution of feature segments in
  music
• Using TF-IDF
• Consider the importance of feature segments in
  the music

13
My thought

• How to divide the frequency domain into n
  sections and decide value of n?
• Why merge the same feature segment?
• Calculating Query similarity
• Precision? Cost?
• Feature work
• Using Run-Length Encoding (RLE) sequences
• Example pattern p AAEEEBBBB
• After RLE p A2E3B4
• It have problem that multi features.