1 Audio Feature Extraction

The following points describe methods to extract various different features from audio files. In most cases only one of these methods is needed.

1.1 Matlab Audio Feature Extraction

•  Rhythm Patterns, SSD, RH
  

1.2 M2K 

• M2K Release Page: Installation, Manuals, Class Overview, etc.
• M2K for MIREX Instructions

1.3 Marsyas 

The vector files with the extracted features can be used as input vectors for the SOM-Toolbox or for classification.

2 SOM-Toolbox

• Download and documentation are available here.
• Our feature extractors above typically write SOMlib Input Vector files, the file format is explained under this link.

 3 Classification

3.1 Weka - Machine Learning Java Tool

• Download Weka. It is implemented in Java, thus platform independent. (On some Linux distributions there might appear problems with the window manager and/or Java 1.5, in that case download the file LookAndFeel.props and place it in your home directory.)
• For Weka you need to convert SomLib vector files into the Weka .arff format. Download the conversion script from SOMLib to .arff format: (right click - Save target as)
  somlib2arff_wclass.pl
  (Perl must be installed to execute the script)
• For classification (and thus for the above script) you need a groundtruth file, where each file to classify is labeled with a class.
  Create a groundtruth file as an Ascii file with a comma-separated list of files and their assigned class:
  <filename>,<class>
  If your files are stored in folders, whose names represent the class labels, (in Linux) you can use commands similar to these in order to create a groundtruth file:
  find . -name *.mp3 > filelist.txt
  cut -d / -f 2 filelist.txt > classes.txt
  paste filelist.txt classes.txt > filelist_wclasses.txt
• Then use the script to create an .arff file:
  somlib2arff_wclass.pl <vector_file> filelist_wclasses.txt
• Follow the Weka-Explorer Instructions for classification.