We have provided you with the data and some initial tools to get you started.

The data folder consists of 10 folders corresponding to each of the 10 genres in the dataset:

1. Blues
2. Classical
3. Country
4. Disco
5. Hiphop
6. Jazz
7. Metal
8. Pop
9. Reggae
10. Rock

Each folder contains the data for 100 songs from the genre in .mat (Matlab) format. We have provided scipts to load the data into Matlab.

The audio files for each song be downloaded from http://opihi.cs.uvic.ca/sound/genres.tar.gz This file is quite large and may take about 10 minutes to fully download. Once unpacked, this folder is organized exactly the same as the data folder.

The data associated with each song is represented as a struct. We provide a script to load the data into matlab, a script to compile all the songs in a single struct, and a script to create CSV files for each song. The .mat files can also be loaded directly into Python.

Each struct contains the following features:

• file_name: 'blues.00001.au'
• class_name: 'blues'
• class: 1
• eng: [1x1198 double]
• mfc: [32x1198 double]
• chroma: [12x1198 double]
• t: [1x1198 double]
• keystrength: [12x1198 double]
• brightness: [1x1198 double]
• zerocross: [1x1198 double]
• roughness: [1x1198 double]
• inharmonic: [1x1198 double]
• key: 8
• tempo: 69.1125

See usicGenreClassificationInformation.pdf in this folder for a description of the features.

We have included some matlab tools to help get started with the data.

• collectfeature.m
  • Extract a feature for all the songs
• demo_fv.m
  • Example use of the Fisher Vector function
• inferqgram.m
  • needed for the exemplar
• loadAll.m
  • Loads files in the data folder into a single structure in Matlab.
• MAT2CSV.py
  • Writes arrays stored within a .mat file to .csv files

Matlab is avaliable free to all students through OIT. Go here: http://www.princeton.edu/software/licenses/software/matlab/ for installation instructions. For any problems with installation, please contact the OIT helpdesk. Matlab is also installed on all machines in campus computer labs.

mfcc and mfc are used interchangeably

To perform the Exemplar and Fisher Vector feature extraction you will need to:

1. Download vlfeat package for MATLAB from http://www.vlfeat.org/download.html
2. Install it in MATLAB using the following instructions: http://www.vlfeat.org/install-matlab.html
3. Refer to demo_fv.m as an example on how to compute Fisher vectors with exemplars
4. To compute FV on an exemplar of size 5+ you will probably need 16GB+ memory

%Change this to the directory containing your data folder

dirn = '/Users/Gazelle/Documents/voxResources';

%intitialize the FV toolbox - you will need to change the filepath appropriately

run('/Users/Gazelle/Documents/MATLAB/vlfeat-0.9.20/toolbox/vl_setup')

%add tools path - you will need to change the filepath appropriately

addpath(genpath('/Users/Gazelle/Documents/voxResources/tools'))

%load all songs into a single struct

[DAT, LB, FNS] = loadAll(dirn);

%extract the MFCC feature

mfcc = cell(1,1000);

for i = 1:length(DAT)

	mfcc{i} = DAT{i}.mfc;

end

%create the structure used as input into the demo_fv

GENDATA.data = mfcc;

GENDATA.class = LB;

GENDATA.classnames = {'Blues', 'Classical', 'Country', 'Disco', 'Hiphop',...

	'Jazz', 'Metal', 'Pop', 'Reggae', 'Rock'};


%run fisher vector

FV = demo_fv(GENDATA, 3, 3)