Y.-C. Lin, Y.-H. Yang, H.-H. Chen, "Exploiting Genre for Music Emotion Classification," Proc. ICME, 2009."

Abstract

Genre and emotion have been applied to content-based music retrieval and organization; however, the intrinsic correlation between them has not been explored. In this paper we present a statistical association analysis to examine such intrinsic correlation and propose a two-layer scheme that exploits the correlation for emotion classification. Significant improvement of classification accuracy over the traditional single-layer scheme is obtained. [Full Text]

Classifiers

In this paper, two different classifiers are used for different purposes. The first one is a multi-class classifier, which is used for genre classificaiton. LIBSVM is adopted in this part. The second one is a multi-label classifier, which is used for emotion classification. A simple extension of LIBSVM is adopted, see the section of "Binary Approach" in LIBSVM tools. The parameter of both classifiers are the default values.

Dataset

The adopted dataset is composed of two parts: labels and audio files. The labels are obtained from All Music Guide, and the audio files are obtained from Youtube. Due to the copyright issue, we only release the 436 features of the audio files extracted by Marsyas.

Since I collect more audio files after submitting the paper, the total number of songs is 2513 not 1353. The dataset can be downloaded below, which contains six files. The description of each file is as below:

  1. X.mat: 436 audio features.
  2. Yorigin.mat: The genre labels and the original emotion labels.
    • Ygenre: 2513*1 matrix, the value of each entry is 1-6, which denote different genres
    • Ymood: 2513*168 matrix, the value of each entry is 0 or 1, where 1 means associating with this emotion, and vice versa.
    • filenames: the adopted files.
  3. Y.mat: the genre labels and the reduced emotion labels.
    • Ygenre: 2513*1 matrix, the value of each entry is 1-6, which denote different genres
    • Ymood: 2513*12 matrix, the value of each entry is 0 or 1, where 1 means associating with this emotion, and vice versa.
    • filenames: the adopted files.
  4. reducedMood.mat: the mapping between original and reduced emotion labels
  5. genres: the adopted genres, the order is as same as the order of the index used in Yorigin.mat
  6. moods: the adopted moods, the order is as same as the order of the index used in Yorigin.mat

dataset

Contact

Any feedbacks or comments are welcomed!
vagante@gmail.com
http://mpac.ee.ntu.edu.tw/~vagante/