O'Reilly logo

Music Emotion Recognition by Homer H. Chen, Yi-Hsuan Yang

Stay ahead with the world's most comprehensive technology and business learning platform.

With Safari, you learn the way you learn best. Get unlimited access to videos, live online training, learning paths, books, tutorials, and more.

Start Free Trial

No credit card required

Saunder January 24, 2011 10:39 book
7
Personalized MER
and Groupwise MER
While Chapter 6 describes the fuzzy approach that intends to address the subjectivity
issue of categorical MER, the following three chapters describe approaches that are
developed to address the subjectivity issue of dimensional MER. We begin with two
intuitive methods proposed in [366] for personalization in this chapter: a personalized
MER scheme that trains a personalized model for emotion prediction using a user’s
feedback and a groupwise MER scheme that trains a number of groupwise regressors
for users possessing different personal factors. Although these two methods have
limited performance, they provide useful insights into the subjectivity issue of MER
and inspire the development of more effective methods.
7.1 Motivation
As described in Chapter 4, regression algorithms can be employed in a dimensional
MER system to train the computational model for predicting the valence and arousal
(VA) values. Typically, a general regressor is trained based upon the average emotion
ratings of music pieces from the subjective test. After training, the general regressor is
applied to predict the emotion values of input music pieces. This approach, however,
neglects the fact that human perception of emotion is by nature subjective and
different people can perceive differently in a music piece. Even though the accuracy
of emotion recognition reaches 100%, it only means that the computational model
accurately predicts only the average emotion perception for a music piece, not the
emotion perception of an individual user for the music piece.
119
Saunder January 24, 2011 10:39 book
120 Music Emotion Recognition
Figure 7.1 shows the annotations of the 60 music pieces used in Chapters 5,
7–9. In this figure, each circle corresponds to the annotation of a subject for a music
piece over the 2D valence-arousal emotion plane. Evidently, simply assigning one
emotion value to a music piece in a deterministic manner does not perform well in
practice because the emotion perception varies greatly from person to person.
To deal with this subjectivity issue, a reasonable approach is to build an MER
system that makes different emotion predictions for different users. This chapter
describes two methods to achieve this:
Personalized MER (PMER): One of the most intuitive methods to resolve
the subjectivity issue is to personalize the MER system. This can be done by
asking a user to explicitly annotate his/her emotion perception of a number
of music pieces and then using these annotations as ground truth to train a
personalized regressor. For the specific user, the personalized regressor should
perform better than the general regressor.subjectivity does play an important
role.
Groupwise MER (GWMER): Another intuitive method groups users accord-
ing to personal factors such as demographic properties, music experience, and
123456
7891011 12
13 14 15 16 17 18
19 20 21 22 23 24
25 26 27 28 29 30
Figure 7.1 The annotations of the 60 songs used in Chapters 5, 7–9.
Saunder January 24, 2011 10:39 book
Personalized MER and Groupwise MER 121
31 35343332 36
37 41403938 42
43 47464544 48
49 53525150 54
55 59585756 60
Figure 7.1 (Continued.)
personality and then trains a groupwise regressor for each user group. One can
interpret the prediction accuracy of a groupwise regressor as the importance
of the corresponding personal factor. For example, if the prediction accuracy
is significantly improved when different regressors are used for people of dif-
ferent genders, it implies that the emotion perception of music is significantly
different between male and female and such differentiation is useful.
This chapter provides the details of these two methods and presents an empirical
performance comparison of the two methods with the general (baseline) method.
7.2 Personalized MER
Though it is great to develop a general MER system that performs equally well for
each user, it may be unnecessary. As pointed out in [252], most of the time it is only
necessary that one’s personal computer is able to recognize his or her emotion. As
emotion perception is by nature subjective, it should be beneficial to personalize the
MER system. Personalized MER (PMER) explores whether the prediction accuracy
for each individual is significantly improved by personalization.

With Safari, you learn the way you learn best. Get unlimited access to videos, live online training, learning paths, books, interactive tutorials, and more.

Start Free Trial

No credit card required