Dataset Description


The dataset consists of two parts:
(1) The campus subset. It is a 240-song subset annotated by 22 subjects recruited from the National Taiwan University and the Academia Sinica.
(2) The Amazon Mechanical Turk (AMT) subset. This subset contains annotations of all the 1608 songs provided by the 643 subjects using AMT Each song receives a total of 15 emotion annotations from each subject in this subset.

For more details about data collection, please refer to our ICASSP 2015 paper.

File Details

The dataset is delivered as a zip file which contains a MATLAB MAT-file named "AMG1608.mat", the README, and a text file named "feature_index.txt" showing the index of each feature set.
AMG1608.mat has three variables, they are:
- feat_cell
- song_info
- song_label
The descriptions of each variable are given in what follows.

This is a 1608-by-1 cell matrix that contains the frame-level acoustic feature vectors of the songs. Each cell is the collection of frame-level feature vectors of a song. The collection of feature vectors is a D-by-F by matrix, where D denotes the feature dimension and F denotes the frame numbers. Specifically, D=72 and F=1199 in this dataset.

This is a 1608-by-2 cell matrix. Each element in the first column is the artist name of each song and each element in the second column is the song title of the song.

This is a 1608-by-665-by-2 cubical matrix that contains the VA annotations of the songs. The three dimensions of this matrix corresponds to the song index, user index, and emotion index, respectively. In particular, the affective valence corresponds to the emotion index=1 and affective arousal corresponds to the emotion index=2.

The campus subset can be selected in MATLAB by the following indexing:

    campus_subset = song_label(1:240, 644:end, :);

On the other hand, the AMT subset can be selected by

    amt_subset = song_label(:, 1:643, :);