Post on 09-Jan-2016
description
Musical Genre Categorization Using Support Vector Machines
Shu Wang
Outline• Motivation• Dataset• Feature Extraction• Automatic Classification • Conclusion
Motivation• Music Information Retrieval
http://www.flickr.com/photos/elbewerk/2845839180/lightbox/ Music Genres
Dataset • GTZAN Genre Collection
• 10 Genres• 30 Seconds Audio Waveform• 1000 Tracks
Dataset: http://marsyas.info/download/data_sets/
Feature Extraction• Features Selection (38 Features)
• Time Domain Zero Crossings• Mel-Frequency Cepstral Coefficients• ….
• Tool• MIRtoolbox
https://www.jyu.fi/hum/laitokset/musiikki/en/research/coe/materials/mirtoolbox
Automatic Classification • Approach
• K-Nearest Neighbors• Support Vector Machine• KNN-SVM Method
Automatic Classification • Difficulty
• Multiclass Classification Problem
• Approach• One versus Rest
• Con: Unbalanced Training Data and Lower Sensitivity and Specificity
• One versus One & Classifier of Classifiers
Training Process
• Each Classifier has high Classification Rate.
Classifier #1 Classifier #2 Classifier #45… …
Blue&Classical Blue&Country Reggae&Rock
Training Process
Testing Process• Combination Rules
• Voting
Classifier #1 Classifier #2 Classifier #45… …
Combination Rules
Testing Features
Final Output
K-Nearest Neighbors• Correct Classification Rate
• 0.6400
• Confusion Matrix36 0 4 2 3
1 1 1 23
0 42 0 0 02 0 0 01
4 3 36 5 00 5 9 613
4 0 1 34 20 2 14 15
1 0 0 2 360 2 1 83
1 4 2 0 046 3 0 24
0 0 2 1 00 36 1 13
0 0 1 3 50 1 17 73
2 0 0 0 40 0 3 220
2 1 4 3 01 0 4 115
K-Nearest Neighbors• Average Correct Classification Rate
• 0.6856
Support Vector Machine• Correct Classification Rate
• 0.6900
• Confusion Matrix35 3 1 1 0
2 2 1 59
0 36 0 1 01 0 0 01
3 2 32 3 02 2 0 54
1 0 4 36 40 2 5 82
1 0 0 0 390 0 1 20
0 7 0 0 041 1 0 10
2 0 1 0 11 36 0 01
0 0 2 5 50 0 40 38
1 1 3 1 10 0 2 261
7 1 7 3 03 7 1 024
Support Vector Machine• Average Correct Classification Rate
• 0.6526
KNN & SVM• Correct Classification Rate
• 0.7100
• Confusion Matrix 40 0 2 2 4 3 1 0 6 1 0 45 0 0 0 3 0 0 0 1 4 1 39 4 0 0 1 4 1 8 1 0 0 30 1 0 3 5 2 2 0 0 0 0 37 0 0 2 13 2 0 2 1 0 0 42 2 0 1 0 2 0 2 1 1 1 41 0 0 7 1 1 1 5 6 0 0 34 4 0 1 0 1 3 1 0 0 1 20 2 1 1 4 5 0 1 2 4 3 27
KNN & SVM• Average Correct Classification Rate
• 0.6928
Conclusion• We achieve over 65% Correct Classification
Rate in this Multiclass Classification Problem
• KNN and SVM method based on One versus One is a promising way to solve the Automatic Genres Classification Problem