The Acoustic Emotion Gaussians (AEG) Model

A Comprehensive Tool for Modeling the Affective Content of Music, Music Emotion Recognition, and Personalization

Authors: Ju-Chiang Wang, Yi-Hsuan Yang, and Hsin-Min Wang
Affiliation: Institute of Information Science, Academia Sinica, Taiwan

Corresponding Email: asriver_DOT_wang_AT_gmail_DOT_com

 

Articles & Credits

Download

We use the DEAP dataset to demonstrate the training and prediction of this AEG Tools. We also use it to demonstrate the leave-one-out or 3-fold cross-validation train-test scenarios.

Download the DEAP dataset: Containing the extracted MFCC features and emotion labels of 120 clips of the DEAP dataset. Please unzip it into the same path of the AEG Tools.

This software requires MATLAB R2007a or later, and supports "MATLABPOOL" for parallelizing the computation. The code of AEG Tools is distributed under the GNU GPL license. See LICENSE for details or http://www.gnu.org/licenses/gpl-3.0.txt

Download the AEG Tools 1.3: Containing the source codes and demonstration scripts. Released in December 2013.

Download the AEG Tools 1.5: Containing the source codes and demonstration scripts. Released in January 2015. This version is more efficient and well-organized. It saves the learned affective GMM at each specified iteration. Different iterations may result in slightly different performance results.

 

Model Visualizations


Figure 1. Visualization of the MFCCs codeword histogram on DEAP.





Figure 2. Visualization of the 2-D affective GMM on DEAP.


Figure 3. Visualization of 2-D personalized affective GMM for the first user in DEAP.





Figure 4. Visualization of the 3-D affective GMM on DEAP.


Figure 5. Visualization of 3-D personalized affective GMM for the first user in DEAP.