Mining For Rock

1
Mining For Rock

• Analyzing WAV music files
• By
• Aaron Morey
• Ryan Mitchell

Prepared for Dr. Craig Struble
2
Overview

• Abstract
• Problem
• Development Process
• Results
• Initial Test Data
• Additional analysis
• Potential improvements
• Conclusion

3
Abstract

• Digital music is becoming more popular
• Easy to have massive collections of music with no
  organization
• iTunes includes genre, type, beats per minute,
  etc.
• Attempt to classify musical by genre using
  digital WAV files
• Genres to be classified
• Classic Rock
• Alternative
• Rap
• Techno

4
Abstract

• Different genres are visually distinct

Classic Rock More Than a Feeling by Boston
5
Abstract

• Different genres are visually distinct

Rap Pass the Dutch by Missy Elliot
6
Abstract

• Different genres are visually distinct

Techno Going Crazy by Sven-R-G
7
Abstract

• Different genres are visually distinct

Alternative Talk by Coldplay
8
Problem

• WAV file format
• Contents
• Header data
• Audio data

9
Problem

• Can be viewed in TextPad
• Numbers represent amplitude of sound wave
• Songs may have 40 million lines

10
Development Process

• Needed a program to read WAV files byte by byte
• Decided to develop in Java
• Using Java objects including File,
  FileInputStream and FileWriter
• Created two new classes
• WAVSong
• WAVReader

11
Development Process

• Data

12
Development Process

• Data

13
Results

• Stages of Test Data
• Progress Report
• Inferior audio quality
• Fewer samples
• Currently Music from CDs
• Recall
• ZeroR 18
• ComplementNaiveBayes 55

14
Results

• Percentage of correct classification on original
  dataset
• ZeroR 29
• JRip 46
• ComplementNaiveBayes 45
• OneR 47
• RandomForest 45

15
Results

• Percentage of correct classification with
  additional data
• ZeroR 34
• JRip 51 (5 Folds)
• ComplementNaiveBayes 44
• OneR 40
• RandomForest 49 (2 seeds)
• Changes in correct classifications
• Overfitting data

16
Results

• Percentage of correct classification with beats
  per minute
• ZeroR 34
• JRip 49 (5 Folds)
• ComplementNaiveBayes 35
• OneR 37
• RandomForest 52
• Changes in correct classifications
• Beats per minute varies widely

17
Potential Improvements

• Use FFT (Fast Fourier transform)
• Java libraries exist
• Too complex to implement in short time
• Header Data
• Read by two-byte sets

18
Sources

• Rauber, Andreas, and Markus Früwirth.
  "Automatically Analyzing and Organizing Music
  Archives." Lecture Notes in Computer Science
  (2004). 28 Mar. 2007 content/4jk8r5mp5gmuj7lw/.
• Maloney, R. Shawn, and Michael Paolisso. "What
  Can Digital Audio Data Do for You?" Field
  Methods, Vol. 13, No. 1, 88-96 (2001) (2001). 28
  Mar. 2007 ract/13/1/88.
• Wilson, Scott. "WAVE PCM Soundfile Format."
  Center for Computer Research in Music and
  Acoustics. 20 Jan. 2003. Stanford University. 28
  Mar. 2007 rojects/WaveFormat/.

19
QUESTIONS?