Some Issues on Detecting Emotions in Music

Piotr Synak and Alicja Wieczorkowska

Polish-Japanese Institute of Information Technology,
Koszykowa 86, 02-008 Warsaw, Poland
{synak, alicja}@pjwstk.edu.pl

Abstract. Investigating subjective values of audio data is both interesting and pleasant topic for research, gaining attention and popularity
among researchers recently. We focus on automatic detection of emotions
in songs/audio les, using features based on spectral contents. The data
set, containing a few hundreds of music pieces, was used in experiments.
The emotions are grouped into 13 or 6 classes. We compare our results
with tests on human subjects. One of the main conclusions is that multilabel classication is required.
Keywords: Music information retrieval, sound analysis.

Introduction

Automatic recognition of emotions in music is a dicult task because of many

reasons. First of all, there is no any universal way or any standard of describing
sound les. Several kind of descriptors can be generated from sounds without
any warranty that they reect any emotions. Moreover, especially in the case
of emotions, any classication (also subjective one) can be ambiguous every
subject may classify emotions in a little bit dierent way. However, for listeners
from similar cultural background, one may expect to obtain similar classication.
Therefore, we considered this topic worth investigations.
We present some initial experiments performed on a database of 870 sound
les classied to 13 or 6 classes of emotions. To describe the les we used a
number of spectral descriptors. In the paper, we discuss the obtained results and
draw the conclusions how to detect the emotions better.

Data Parametrization

Automatic parametrization of audio data for classication purposes is hard because of ambiguity of labeling and subjectivity of description. However, since a
piece of music evokes similar emotions in listeners representing the same cultural
background, it seems to be possible to obtain parametrization that can be used
for the purpose of extracting emotions. Our goal was to check how numerical
parameters work for classication purposes, how good or low is classication
accuracy, and how it is comparable with human performance.
ezak et al. (Eds.): RSFDGrC 2005, LNAI 3642, pp. 314322, 2005.
D. Sl


c Springer-Verlag Berlin Heidelberg 2005

Some Issues on Detecting Emotions in Music

315

Objective descriptors of audio signal characterize basic properties of the investigated sounds, such as loudness, duration, pitch, and more advanced properties, describing frequency contents and its changes over time. Some descriptors
come from speech processing and include prosodic and quality features, such as
phonation type, articulation manned etc. [12]. Such features can be applied to
detection of emotions in speech signal, but not all of them can be applied to
music signals, which require other descriptors. Features applied to music signal include structure of the spectrum - timbral features, time domain features,
time-frequency description, and higher-level features, such as rhythmic content
features [7], [9], [13], [14].
When parameterizing music sounds for emotion classication, we assumed
that emotions depend, to some extend, on harmony and rhythm. Since we deal
with audio, not MIDI les, our parametrization is based on spectral contents
(chords and timbre). Western music, recorded stereo with 44100 Hz sampling
frequency and 16-bit resolution was used as audio samples. We applied long
analyzing frame, 32768 samples taken from the left channel, in order obtain more
precise spectral bins, and to describe longer time fragment. Hanning window was
applied, and spectral components calculated up to 12 kHz and no more than 100
partials, since higher harmonics did not contribute signicantly to the spectrum.
The following set of 29 audio descriptors was calculated for our analysis
window [14]:
F requency: dominating fundamental frequency of the sound
Level: maximal level of sound in the analyzed frame
T ristimulus1, 2, 3: Tristimulus parameters calculated for F requency, given
by [10]:
A2
(1)
T ristimulus1 =
N 1
2
n=1 An


2
n=2,3,4 An
(2)
T ristimulus2 =
N
2
n=1 An

N
A2n
T ristimulus3 =
n=5
(3)
N
2
n=1 An
where An denotes the amplitude of the nth harmonic, N is the number of
harmonics available in spectrum, M =
N/2 and L =
N/2 + 1
EvenHarm and OddHarm: Contents of even and odd harmonics in the
spectrum, dened as


M
2
k=1 A2k
EvenHarm = 

(4)
N
2
n=1 An


L
2
k=2 A2k1
(5)
OddHarm = 

N
2
A
n=1 n

316

P. Synak and A. Wieczorkowska

Brightness: brightness of sound - gravity center of the spectrum, dened as

N
Brightness =
n=1
N

n An

n=1

An

(6)

Irregularity: irregularity of spectrum, dened as [5], [6]


Irregularity = log 20




Ak


log 

3

Ak1 Ak Ak+1 

N
1 

k=2

(7)

F requency1, Ratio1, ..., 9: for these parameters, 10 most prominent peaks

in the spectrum are found. The lowest frequency within this set is chosen
as F requency1, and proportions of other frequencies to the lowest one are
denoted as Ratio1, ..., 9
Amplitude1, Ratio1, ..., 9: the amplitude of F requency1 in decibel scale,
and dierences in decibels between peaks corresponding to Ratio1, ..., 9 and
Amplitude1. These parameters describe relative strength of the notes in the
music chord.

Experiment Setup

Investigations on extracting emotions from music data were performed on a

database of 870 audio samples. The samples represented 30 seconds long excerpts from songs and classic music pieces. This database was created by Dr.
Rory A. Lewis from the University of North Carolina at Charlotte. Therefore,
all audio le were labeled with information about emotions by a single subject.
The pieces were recorded in MP3 format and next converted to au/snd format
for parametrization purposes. Sampling frequency 44100 Hz was chosen. Parametrization was performed for 32768 samples (215 ) frame length. The data set
is divided into the following 13 classes, covering wide range of emotions [7]:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.

frustrated,
bluesy, melancholy,
longing, pathetic,
cheerful, gay, happy,
dark, depressing,
delicate, graceful,
dramatic, emphatic,
dreamy, leisurely,
agitated, exciting, enthusiastic,
fanciful, light,
mysterious, spooky,
passionate,
sacred, spiritual.

Some Issues on Detecting Emotions in Music

Class
No. of objects
Agitated
74
Bluesy
66
Dark
31
Dramatic
101
Dreamy
46
Fanciful
38
Frustrated
152

317

Class
No. of objects
Graceful
45
Happy
36
Passionate
40
Pathetic
155
Sacred
11
Spooky
77

Fig. 1. Representation of classes in the 870-element database

Number of samples in each class is shown in Figure 1.

Some classes are underrepresented, whereas others are overrepresented in
comparison with the average number of objects in a single class. Moreover, labeling of classes is dicult in some cases, since the same piece may evoke various
emotions. Therefore, we decided to join the data into 6 superclasses as follows
(see [7]):
1.
2.
3.
4.
5.
6.

happy and fanciful,

graceful and dreamy,
pathetic and passionate,
dramatic, agitated, and frustrated,
sacred and spooky,
dark and bluesy.

The classication experiments were performed using k-NN algorithm, with k

varying within range 1..20, and the best k in each experiment was chosen. We
decided to use k-NN because in the rst experiments it outperformed classiers
of other types. For training purposes, 20% of the data set was removed and then
used as test data after nishing training; this procedure was repeated 5 times
(i.e., standard CV-5 procedure was applied). Next, the results were averaged. In
order to compare results with Li and Ogihara [7], we performed experiments for
each class separately, recognizing them in a binary way - one class against the
rest of the data. The binary classication can be a good basis for construction
of a general classier, based on a set of binary classiers [1].

Results

The experiments described in the previous section were rst performed on a

smaller data set, containing 303 objects, as presented in Figure 2. These experiments yielded results presented in Figure 3.
The results can be well compared with the results obtained by Li and Ogihara
[7]. They obtained accuracy ranging from 51% to 80% for various classes and
30-element feature vector, with use of 50% of data for training and the remaining
50% of the data set, consisting of 599 audio les, also labeled by a single subject
into the same 3 classes, and then into 6 classes, as described in section 3. We

318

P. Synak and A. Wieczorkowska

Class
No. of objects
Agitated
16
Bluesy
18
Dark
6
Dramatic
88
Dreamy
20
Fanciful
34
Frustrated
17

Class
No. of objects
Graceful
14
Happy
24
Passionate
18
Pathetic
32
Sacred
17
Spooky
7

Fig. 2. Representation of classes in the collection of 303 musical recordings for the
research on automatic classifying emotions

Class
No. of objects k-NN Correctness
1. happy, fanciful
57
k=11
81.33%
2. graceful, dreamy
34
k=5
88.67%
3. pathetic, passionate
49
k=9
83.67%
4. dramatic, agitated, frustrated
117
k=7
62.67%
5. sacred, spooky
23
k=7
92.33%
6. dark, bluesy,
23
k=5
92.33%
Fig. 3. Results of automatic classication of emotions for the 303-element database
using k-NN

Class
No. of objects Correctness
1. happy, fanciful
74
95.97%
2. graceful, dreamy
91
89.77%
3. pathetic, passionate
195
71.72%
4. dramatic, agitated, frustrated
327
64.02%
5. sacred, spooky
88
89.88%
6. dark, bluesy,
97
88.80%
Fig. 4. Results of automatic classication of emotions for the 870-element database

also performed experiments for the same 6 classes, using k-NN classier, i.e.,
examining all 6 classes in parallel. These experiments yielded 37% correctness
(and 23.05% for 13 classes), suggesting that further work was needed. Since we
suspected that uneven number of objects in classes and not too big data set could
hinder classication, the 870-element data set was used in further experiments.
The results of experiments with binary classication performed on the full
data set, containing 870 audio les, are presented in Figure 4. The best results
of experiments were obtained in k-NN for k = 13. As we can see, the results
have been even improved comparing to the small data set. However, general
classication for all classes examined in parallel was still low, comparable with

Some Issues on Detecting Emotions in Music

319

results for 303-element data set, since we obtained 20.12% accuracy for 13 classes
and 37.47% for 6 classes.
Because of the low level of accuracy in general classication, we decided to
compare the results with human performance. Two other subjects with musical
background were asked to classify a test set of 39 samples, i.e., 3 samples for
each class. The results convinced us that the diculty is not just in the parametrization or method of classication, since the correctness of assessment yielded
24.24% and 33.33%, diering essentially on particular samples. This experiment
suggests that multi-class labeling by a few subjects may be needed, since various
listeners may perceive various emotions while listening to the same le, even if
they represent the same cultural and musical background.

Multi-class Labeling

In our experiments we used a database of music les collected by a professional

musician acting as an expert. Every le was labeled and classied to exactly one
class representing particular kind of emotion. One of the rst remarks of the
expert was that in several cases it is impossible to classify a song to exactly one
class. First of all, the nature of the song and the melody can be labeled by more
than one adjective. Secondly, labeling is very subjective and dierent people may
use various associations. This is because the perception of sounds by humans is
not uniform. The perception can be dominated by dierent factors, e.g., by
particular instrument or by vocal, and thus, dierent labels can be attached to
the same piece of sound.
The results of our initial experiments (i.e., with 13 decision classes) one may
interpret as not satisfactory. One reason of low results is that the set of used
descriptors is relatively small. However, we claim that the most important factor
is that the les were initially classied to single classes only. To conrm this we
conducted the following experiment. We asked another two musicians (female
and male) to classify the sound les to the same categories as the initial labeling.
As we stated in the previous section, the results were very surprising in that the
quality of recognition by a human was worse than one obtained by k-NN.
From the discussions with the experts it follows that the main diculty while
performing the classication was that they had to choose one class only, whilst
in most cases they found at least two class labels appropriate. Therefore, we
suggest to use multi-class labeling, i.e., to allow labeling each piece of sound
with any number of labels.
The data that can be classied to more than one class are known in the
literature as multi-label data [2,8,4]. This kind of data is often being analyzed
in text mining and scene classication, where text documents or pictures may
have been attached several labels describing their contents.
There are several problems related to multi-label data analysis, including:
selecting training model with multi-label data, using testing criteria, and evaluating multi-label classication results.

320

5.1

P. Synak and A. Wieczorkowska

Training Models

One of the basic questions of training phase of classiers induction is how to

use training examples with multiple labels? There are a few models commonly
used.
The simplest model (MODEL-s) assumes labeling of data by using single
label the one which is most likely.
MODEL-i assumes ignoring all the cases with more than one label. That
means that there can no data to be used in the training phase if there are no
data with single label.
In MODEL-n there are created new classes for each combination of labels
occurring in the training sample. The main problem of this model is that the
number of classes easily becomes very large, especially when we consider not
only two, but three and more labels attached to one sample. Therefore, the data
become very sparse, and, as result of that, several classes can have very few
training samples.
The most ecient model seems to be MODEL-x, cross-training, where samples with many labels are used as positive examples, and not as negative examples, for each class corresponding to the labels.
5.2

Testing Criteria

We assume that we build models for each base class only, and not for combination
of classes (MODEL-n) because of sparseness of data as discussed above. As an
exemplary classier we use Support Vector Machines (SVM) [3] as they are
recognized to give very good results in text and scene classication, i.e., in multilabel data.
Now, let us see how can we obtain multiple labels from the outputs of each
of the models. In standard 2-class SVM the positive (negative) output of a SVM
for a testing object means that it is a positive (negative) example. In the case
of multi-class problems there are several SVMs built one for each class. The
highest positive output of SVMs determines the class of a testing object. However, it can happen that no SVM gives positive output. This approach can be
extended to multi-label classication.
Let us consider the following three testing (labeling) criteria.
P-criterion labels the testing object with all classes corresponding to positive
output of SVM. If no output is positive than the object is unlabeled.
T-criterion works similarly to P-criterion, however, if no output is positive
than the top value is used for labeling.
C-criterion evaluates top values that are close each other no matter whether
they are positive or negative.
5.3

Evaluating Classication Results

Evaluation of results diers from the classical case of single-label classication,

where testing object is classied either correctly or incorrectly. In the case of
multi-label data classication we can have more cases. If all the labels assigned

Some Issues on Detecting Emotions in Music

321

to a testing object are proper then it is classied correctly if all are wrong
then incorrectly. However, what makes it dierent from single-label classication,
only some of the labels can be attached properly this is the case of partial
correctness.
Thus, except standard measures of quality of classication like precision or
accuracy, we need additional ones that take into account also partial correctness.
Some examples of such measures, for example one-error, coverage, and precision,
have been proposed in the literature (see, e.g., [11]). In [2] there are proposed
two methods, -evaluation and base class evaluation of multi-label classier evaluation that make it possible to analyze results of classication in a wide range
of settings.

Conclusions and Future Work

Dicult task of automatic recognition of emotions in music pieces was investigated in our research. The purpose of this investigations was not only testing
how numerical parameters perform in objective description of subjective features, but also assessment of the recognition accuracy, and comparison of results
with human subjects. The obtained accuracy is not high, but is of the same
quality as human assessment. Since humans dier in their opinions regarding
emotions evoked by the same piece, inaccuracies in automatic classication are
not surprising.
In the next experiments we plan to apply multi-class labeling of sounds and
develop the methodology for multi-label data classication.
We also plan to extend the set of descriptors used for sounds parametrization.
In particular, we want to investigate how values of particular parameters change
with time and how it is related to any kind of emotions.

Acknowledgements
This research was partially supported by the National Science Foundation under grant IIS-0414815, by the grant 3 T11C 002 26 from Ministry of Scientic
Research and Information Technology of the Republic of Poland, and by the
Research Center at the Polish-Japanese Institute of Information Technology,
Warsaw, Poland.
The authors express thanks to Dr. Rory A. Lewis from the University of North
Carolina at Charlotte for elaborating the audio database for research purposes.

References
1. Berger, A.: Error-correcting output coding for text classication. IJCAI99: Workshop on machine learning for information ltering. Stockholm, Sweden (1999).
Available at http://www-2.cs.cmu.edu/aberger/pdf/ecoc.pdf
2. Boutell, M., Shen, X., Luo, J., Brown, C.: Multi-label Semantic Scene Classication.
Technical Report, Dept. of Computer Science, U. Rochester, (2003).

322

P. Synak and A. Wieczorkowska

3. Burges, C.J.: A tutorial on support vector machines for pattern recognition. Data
mining and knowledge discovery 2(2) (1998) 121167.
4. Clare, A., King, R.D.: Knowledge Discovery in Multi-label Phenotype Data. Lecture Notes in Computer Science 2168 (2001) 4253.
5. Fujinaga, I., McMillan, K.: Realtime recognition of orchestral instruments. Proceedings of the International Computer Music Conference (2000) 141143
6. Kostek, B., Wieczorkowska, A.: Parametric Representation Of Musical Sounds.
Archives of Acoustics 22, 1 (1997) 326
7. Li, T., Ogihara, M.: Detecting emotion in music. 4th International Conference on
Music Information Retrieval ISMIR, Washington, D.C., and Baltimore, MD (2003).
Available at http://ismir2003.ismir.net/papers/Li.PDF
8. McCallum, A.: Multi-label Text Classication with a Mixture Model Trained by
EM. AAAI99 Workshop on Text Learning, (1999).
9. Peeters, G. Rodet, X.: Automatically selecting signal descriptors for Sound Classication. ICMC 2002 Goteborg, Sweden (2002)
10. Pollard, H. F., Jansson, E. V.: A Tristimulus Method for the Specication of Musical Timbre. Acustica 51 (1982) 162171
11. Schapire, R., Singer, Y.: Boostexter: a boosting-based system for text categorization. Machine Learning 39(2/3) (2000) 135168
12. Tato, R., Santos, R., Kompe, R., Pardo, J. M.: Emotional Space Improves Emotion
Recognition. 7th International Conference on Spoken Language Processing ICSLP
2002, Denver, Colorado (2002).
13. Tzanetakis, G., Cook, P.: Marsyas: A framework for audio analy
sis. Organized Sound 4(3) (2000) 169-U175.
Available at http://www2.cs.cmu.edu/gtzan/work/pubs/organised00gtzan.pdf
14. Wieczorkowska, A., Wroblewski, J., Synak, P., Slezak, D.: Application of temporal
descriptors to musical instrument sound recognition. Journal of Intelligent Information Systems 21(1), Kluwer (2003), 7193