From Content-based Music Emotion Recognition to Emotion Maps of Musical Pieces by Jacek Grekow

Author:Jacek Grekow
Language: eng
Format: epub
Publisher: Springer International Publishing, Cham

7.2 Related Work

Studies on emotion detection in music are mainly based on two popular approaches: categorical or dimensional. In the dimensional approach, emotions are described as numerical values of valence and arousal. The categorical approach describes emotions with a discrete number of classes – affective adjectives. In this chapter, we used the categorical approach.

One of the first papers on categorical emotion detection was a study by Li and Ogihara [57], who trained support vector machines (SVM) to classify music into one of 13 mood categories using a multi-label classification method. A labeled collection consisted of 499 sound files (30-seconds each) from the ambient, classical, fusion, and jazz genres. They used Marsyas to extract the timbral, rhythmic, and pitch features. The achieved accuracy was low, at a level of 45%.

Lu et al. [63] examined emotion detection and emotion tracking using intensity, timbre, and rhythm acoustic features. Emotion categories corresponded to the four quadrants on Thayer’s two-dimensional (Energy-Stress) model [103]. To train, Gaussian Mixture Models were used on a set of 800 classical music clips (20 s each). The system of emotion detection achieved an average accuracy of 86%. In addition to emotion detection, emotion tracking through a music piece was presented, which divided the music into several segments.

The problem of multi-label classification of emotions in musical recordings was also presented by Wieczorkowska et al. [113]. The data set contained 875 samples with a length of 30 s each. For classification, the k-nearest neighbors (k-nn) algorithm was used.

In the community of Music Information Retrieval Evaluation eXchange (MIREX) for automatic music mood classification, five mood clusters were used for song categorization [43]. The Audio Mood Classification evaluation task was started for the first time in 2007. The ground truth set consisted of 600 clips (30 second each), with 120 in each mood cluster. The five emotion clusters, which were used by MIREX Audio Mood Classification, have not been frequently used in other music emotion detection works. Hu et al. in [44] indicates that the clusters might not be optimal and noticed some semantic overlap.

A popular emotion set used to categorize emotions in music turned out to be a collection consisting of 4 classes: happy, angry, sad, and relaxed. It corresponds to the four quadrants of the two-dimensional valence-arousal plane, which was used by Laurier in [54], where binary classifiers were constructed for each category. A data set of 1000 songs (30 s each) was divided between 4 categories. Classification accuracy was from 84% to 98%, and was obtained for the SVM algorithm with polynomial and linear kernel.

Four emotion classes (happy, angry, sad, relaxed) were also used in the categorical approach by Song et al. in [100]. The collected ground truth data set consisted of 2904 songs that were labeled with one of the four emotions. The highest accuracy, 53%, was achieved for SVM with polynomial kernel. Song et al. explored the relationship between musical features extracted by MIRtoolbox [53] and emotions. They compared the emotion prediction results for four sets of features: dynamic, rhythm, harmony, and spectral.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.