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
6
Fuzzy Classification
of Music Emotion
Emotion perception is intrinsically subjective, and people usually perceive different
emotions in the same song. This subjectivity issue makes the performance evaluation
of an MER system fundamentally difficult because a common agreement on the
classification result is hard to obtain. An MER system that neglects the subjectivity
issue and assigns emotion labels or values to music pieces in a deterministic fashion
cannot perform well in practice, since it is likely that a user is not satisfied with the
prediction result. The following four chapters describe a number of methods that
have been developed to address the subjectivity issue, including techniques for soft
assignment and personalization. We begin with a fuzzy approach that is developed for
categorical MER in this chapter. This approach estimates the likelihood of perceiving
an emotion when listening to a music piece, thereby making the prediction result
less deterministic.
6.1 Motivation
Due to the subjective nature of human perception, classification of the emotion
of music is a challenging problem. Cultural background, age, gender, personality,
training, and so forth can influence the emotion perception [16, 147]. Because
of these factors, classification methods (e.g., [217, 336, 352]) that simply assign
one emotion class to each song in a deterministic manner do not perform well in
practice.
The subjective nature of emotion perception suggests that fuzzy logic is an ap-
propriate mathematical tool for emotion detection [83, 155]. For example, fuzzy
107
Saunder January 24, 2011 10:39 book
108 Music Emotion Recognition
classifiers can be employed to measure the strength of an emotion class in association
with the music piece under classification. Based on the measurement, the user knows
how likely a music piece belongs to an emotion class. In this way, the classification
result is less deterministic and more acceptable to users.
The idea of applying fuzzy classifiers to the categorical approach to MER is first
proposed in [365], which represents one of the first attempts to take the subjective na-
ture of human perception into consideration for MER. This chapter provides the de-
tails of this fuzzy approach. Another approach that utilizes the probabilistic estimate
of support vector machine (SVM) [254] to make soft assignment of classification re-
sult is later proposed in [344]. This prediction result of the probabilistic approach is
the probability distribution of a music piece over the Hevner’s eight emotion classes,
which is called the emotion histogram.Asdescribed in Chapter 9, the idea of making
soft predictions has also been applied to the dimensional approach to MER.
6.2 Fuzzy Classification
Compared with traditional classifiers that only assign one class to a test sample, a
fuzzy classifier assigns a fuzzy vector that indicates the relative strength of each class.
For example, assuming our emotion taxonomy consists of four emotion classes, a
fuzzy vector of [0.10.00.80.1]
indicates a fairly strong emotion strength for the
third class, while [0.10.40.40.1]
shows an ambiguity between the second and
the third classes. The ambiguity that fuzzy vectors carry is very important because
emotion perception is intrinsically subjective.
Many algorithms have been developed for fuzzy classification. Below we describe
two of the most popular algorithms, fuzzy k-nearest neighbor and fuzzy nearest-mean.
6.2.1 Fuzzy k-NN Classifier
The (crisp) k-nearest neighbor (k-NN) classifier is commonly used in pattern recog-
nition [78]. A test sample is assigned to the class that represents the majority of
class labels of the k-nearest neighbors of the test sample, where the distance between
samples is measured in the feature space. However, only a class label is assigned to
the test sample. There is no indication of its strength of membership in that class.
Fuzzy k-NN classifier [165], a combination of fuzzy logic and k-NN classifier,
is designed to solve the above problem. It contains two steps: fuzzy labeling that
computes the fuzzy vectors of training samples (done in model training) and fuzzy
classification that computes the fuzzy vectors of test samples (done in testing).
In fuzzy labeling,wecompute µ
i
, the fuzzy vector of a training sample.
Several methods have been developed to compute the fuzzy vector [115,165]. These
methods can be generally described by the following formula:
µ
ic
=
β +
n
c
k
(1 β), if c = v
n
c
k
(1 β), otherwise
(6.1)

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