Combining Musical and Cultural Features for Intelligent Style Detection

1
Combining Musical and Cultural Features for
Intelligent Style Detection

• Brian Whitman
• Paris Smaragdis
• MIT Media Lab

2
Background

• Music classification by style
• A human concept hard to model.
• Defines subclasses of genres.
• Can be utilized by recommendation engine for
  high-confidence results.

3
Approach

• An automatic style detection system that operate
  on both of
• acoustic content of the audio
• community metadata a vector space of descriptive
  textual terms crawled from the web
• Dataset 5 styles, each with 5 different artists

4
Audio-based Classification

• Form each song into some presentation
• Train a neural network to classify a song

5
Audio-based Classification Result
Heavy Metal Contemporary Country Hardcore Rap Intelligent Dance Music RB
6
Community Metadata-based Classification (1)

• Cultural feature
• Each artist is associated with terms which appear
  on the same web document as the artists name.
• Each term has a score calculated in terms of
  position and frequency of occurrence.

7
Community Metadata-based Classification (2)

• Similarity
• For every 2 artists, calculate an overlap weight,
  which is the summation of every shared term.
• Form a similarity matrix to predict the style of
  each artist

8
Community Metadata-based Classification - Result
Heavy Metal Contemporary Country Hardcore Rap Intelligent Dance Music RB
9
Combined Classification
Heavy Metal Contemporary Country Hardcore Rap Intelligent Dance Music RB
10
Conclusion Future Work

• Combined classification can overcome all the
  problems
• Future development can use a culture ratio to
  alert recommendation engines to use which
  classification method.