See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221558637

Feature Selection in Automatic Music Genre Classification

Conference Paper · December 2008

DOI: 10.1109/ISM.2008.54 · Source: DBLP

CITATIONS READS
37 520

3 authors:

Carlos Silla Alessandro L Koerich

Pontifícia Universidade Católica do Paraná (PUC-PR) École de Technologie Supérieure
80 PUBLICATIONS 2,048 CITATIONS 178 PUBLICATIONS 2,605 CITATIONS

SEE PROFILE SEE PROFILE

Celso A. A. Kaestner
Federal University of Technology - Paraná/Brazil (UTFPR)
65 PUBLICATIONS 1,169 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

End-to end systems based on deep learning for environmental sound classification View project

Aplicação da otimização por colônia de formigas ao problema de múltiplos caixeiros viajantes no atendimento de ordens de serviço nas empresas de distribuição de
energia elétrica View project

All content following this page was uploaded by Celso A. A. Kaestner on 28 July 2014.

The user has requested enhancement of the downloaded file.

 Tenth IEEE International Symposium on Multimedia

Feature Selection in Automatic Music Genre Classification

Carlos N. Silla Jr.

University of Kent
Computing Laboratory
Canterbury, CT2 7NF, Kent, United Kingdom
[email protected]
Alessandro L. Koerich
Pontifical Catholic University of Paraná
R. Imaculada Conceição 1155, 80215-901, Curitiba, Brazil
[email protected]
Celso A. A. Kaestner
Federal University of Technology of Paraná
Av. Sete de Setembro 3165, 80230-901, Curitiba, Brazil
[email protected]

Abstract 1. Introduction

Music genres can be defined as categorical labels cre-

ated by humans in order to identify the style of the music.
The automatic classification of music genres is nowadays
an important task, because music genre is a descriptor that
This paper presents the results of the application of a fea- is largely used to organize large collections of digital mu-
ture selection procedure to an automatic music genre clas- sic [1], [21]. This is specially true in the Internet, which
sification system. The classification system is based on the contains large amounts of multimedia content, and where
use of multiple feature vectors and an ensemble approach, music genre is frequently used in search queries [6], [9].
according to time and space decomposition strategies. Fea- Also, from a pattern recognition perspective, the task of au-
ture vectors are extracted from music segments from the be- tomatic music genre classification poses an interesting re-
ginning, middle and end of the original music signal (time- search problem: music signal, a complex time-variant sig-
decomposition). Despite being music genre classification a nal, is very high dimensional, and music databases can be
multi-class problem, we accomplish the task using a com- very large [2].
bination of binary classifiers, whose results are merged in Most of the current research on music genre classifica-
order to produce the final music genre label (space decom- tion focus on the development of new feature sets and clas-
position). As individual classifiers several machine learning sification methods [10], [11], [14]. On the other hand, few
algorithms were employed: Naı̈ve-Bayes, Decision Trees, works have dealt with feature selection. One of the few ex-
Support Vector Machines and Multi-Layer Perceptron Neu- ceptions is the work of Grimaldi et al. [8] which presents
ral Nets. Experiments were carried out on a novel dataset a new method for feature extraction based on the discrete
called Latin Music Database, which contains 3,227 music wavelet transform; however, no experiments have been per-
pieces categorized in 10 musical genres. The experimen- formed using a standard set of features, like the ones pro-
tal results show that the employed features have different posed by Tzanetakis & Cook [21]. More recently Fiebrink
importance according to the part of the music signal from & Fujinaga [7] have employed a forward feature selec-
where the feature vectors were extracted. Furthermore, the tion (FFS) procedure and the principal component anal-
ensemble approach provides better results than the individ- ysis (PCA) procedure for automatic music classification.
ual segments in most cases. Yaslan and Cataltepe [23] have also used a feature selec-
tion (FS) for music classification using dimensionality re-

978-0-7695-3454-1/08 $25.00 © 2008 IEEE 39

DOI 10.1109/ISM.2008.54

Authorized licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.
duction methods, such as forward (FFS) and backward fea- tracted from three 30-second segments (equivalent to 1,153
ture selection (BFS) and PCA. The results suggest that fea- frames in a MP3 file) from the beginning, middle and end
ture selection, the use of different classifiers, and a subse- parts of the original music. We argue that this procedure is
quent combination of results can improve the music genre adequate for the problem, since it can better treat the time
classification accuracy. Bergstra et al. [2] use the ensem- variation that is usual in music pieces. Also, it allows us to
ble learner AdaBoost which performs the classification it- evaluate if the features extracted from different parts of the
eratively by combining the weighted votes of several weak music have similar discriminative power. Figure 1 illustrate
learners. The procedure shows to be effective in three mu- this process.
sic genre databases, winning the music genre identification
task in the MIREX 2005 (Music Inf. Retrieval EXchange).
The aim of this work it to apply a feature selection pro-
cedure, based on Genetic Algorithms (GA), to multiple fea-
ture vectors extracted from different parts of the music sig-
nal, and analyze the discriminative power of the features
according to the part of the music signal from where they
were extracted, and the impact of the feature selection on
the music genre classification. Another reason for the use
of a GA-based FS, instead of other techniques such as PCA,
is that the GA is a more profitable approach from a musico-
logical perspective, as pointed out in [13]. Figure 1. Time Decomposition Approach
This paper is organized as follows: Section 2 presents the
time/space decomposition strategies used in our automatic
music classification system; Section 3 presents the feature 2.2 Space decomposition
selection procedure; Section 4 describes the dataset used in
the experiments and the results achieved while using feature Despite being music genre classification naturally a
selection over multiple feature vectors. Finally, the conclu- multi-class problem, we accomplish the task using a com-
sions are stated in the last section. bination of binary classifiers, whose results are merged in
order to produce the final music genre labeling. Since dif-
2 Music classification: the time/space decom- ferent features are used for different classes, the procedure
position approach characterize a space decomposition of the feature space,
justified because in this case the classifiers tend to be sim-
Music genre classification can be considered as a three ple and effective [12]. Two main techniques are employed:
step process [2]: (1) the extraction of acoustic features from (a) in the one-against-all (OAA) approach, a classifier is
short frames of the audio signal; (2) the aggregation of the constructed for each class, and all the examples in the re-
features into more abstract segment-level features; and (3) maining classes are considered as negative examples of that
the prediction of the music genre using a classification al- class; (b) in the round-robin (RR) approach, a classifier is
gorithm that uses the segment-level features as input. constructed for each pair of classes, and the examples be-
In this work we employ the MARSYAS framework [21] longing to the other classes are discarded. Figures 2 and 3
for feature extraction; it extracts acoustic features from the illustrate these approaches.
audio frames and aggregate them into music segments. Our For a m-class problem (m music genres) several classi-
music classification system is based on standard supervised fiers are generated: m classifiers in OAA and m(m − 1)/2
machine learning algorithms. However, we employ multi- classifiers in RR. The output of these classifiers are com-
ple feature vectors, obtained from the original music signal bined according to a decision procedure in order to produce
according to time and space decompositions [4], [20], [17]. the final class label.
Therefore several feature vectors and component classifiers
are used in each music part, and a combination procedure 2.3 Feature set
is employed to produce the final class label, according to an
ensemble approach [12]. There is no accepted theory of which features are ade-
quate for music classification tasks [1], [2]. In our work
2.1 Time decomposition we employ the MARSYAS framework for feature extraction
from each music segment. This framework implements the
The music signal is naturally a time varying signal. Time original feature set proposed by Tzanetakis & Cook [21].
decomposition is obtained considering feature vectors ex- The features can be split in three groups: Beat Related,

40

Authorized licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.
 Table 1. Feature vector description
Feature # Description
1 Relative amplitude of the first histogram peak
2 Relative amplitude of the second histogram peak
3 Ratio between the amplitudes of the second peak
and the first peak
4 Period of the first peak in bpm
5 Period of the second peak in bpm
6 Overall histogram sum (beat strength)
7 Spectral centroid mean
8 Spectral rolloff mean
9 Spectral flow mean
Figure 2. One-Against-All Space Decomposi-
10 Zero crossing rate mean
tion Approach 11 Standard deviation for spectral centroid
12 Standard deviation for spectral rolloff
13 Standard deviation for spectral flow
14 Standard deviation for zero crossing rate
15 Low energy
16 1 rt. MFCC mean
17 2 nd. MFCC mean
18 3 rd. MFCC mean
19 4 th. MFCC mean
20 5 th. MFCC mean
21 Standard deviation for 1 rt. MFCC
22 Standard deviation for 2 nd. MFCC
23 Standard deviation for 3 rd. MFCC
24 Standard deviation for 4 th. MFCC
25 Standard deviation for 5 th. MFCC
Figure 3. Round-Robin Space Decomposition 26 The overall sum of the histogram (pitch strength)
Approach 27 Period of the maximum peak of the
unfolded histogram
28 Amplitude of maximum peak of the
Timbral Texture and Pitch Related. The Beat-Related fea- folded histogram
29 Period of the maximum peak of the
tures (features 1 to 6) include the relative amplitudes and the
folded histogram
beats per minute. The Timbral Texture features (features 7
30 Pitch interval between the two most prominent
to 25) account for the means and variance of the spectral peaks of the folded histogram
centroid, rolloff, flux, the time zero domain crossings, the
first 5 Mel Frequency Cepstral coefficients and low energy.
Pitch Related features (features 26 to 30) include the maxi- 2.4 Classification, Combination and Deci-
mum periods and amplitudes of the pitch peaks in the pitch sion
histograms. We note that most of the features are calculated
over time intervals. Standard machine learning algorithms were employed as
A normalization procedure is applied, in order to homog- individual component classifiers. Our approach is homoge-
enize the input data for the classifiers: if maxV and minV neous, that is, the very same classifier is employed in every
are the maximum and minimum values that appears in all music part. In this work we use the following algorithms:
dataset for a given feature, a value V is replaced by newV Decision Trees (J48), k-NN, Naı̈ve-Bayes (NB), a Multi-
using the equation layer Perceptron Neural Network Classifier (MLP) with the
(V − minV ) backpropagation momentum algorithm, and a Support Vec-
newV = tor Machines (SVM) with pairwise classification [15]. All
(maxV − minV )
the experiments were conducted in a framework based on
The final feature vector, outlined at Table 1, is 30- the WEKA Datamining Tool [22].
dimensional (Beat: 6; Timbral Texture: 19; Pitch: 5). For The final classification label is obtained from all the par-
a more detailed description of the features refer to [21] or tial classifications, by using a decision procedure. In our
[18]. case, the combination of the time and space decomposition

41

Authorized licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.
strategies works as follows: (1) one of the space decompo- 4 Experiments
sition approaches (RR or OAA) is applied to all three seg-
ments of the time decomposition approach (i.e. beginning, This section presents the experiments and the results
middle and end); (2) a local decision considering the class achieved on music genre classification and feature selection.
of the individual segment is made based on the underlying The main goal is to evaluate if the features extracted from
space decomposition approach: the majority vote for the RR different origins in the audio signal have similar discrimi-
and rules based on the a posteriori probability given by the native power for music genre classification. Another goal
specific classifier of each case for the OAA; (3) the decision is to verify if the ensemble-based method provides better
concerning the final music genre of the song is made based results than the classifiers taking into account features ex-
on the majority vote of the predicted genres from the three tracted from single segments.
individual segments. We employ the new Latin Music Database 1 [19], [18]
which contains 3,227 MP3 music pieces from 10 different
3 Feature Selection Latin genres, originated from music pieces of 501 artists. In
this database music genre assignment was manually made
The task of feature selection (FS) consists in choosing by a group of human experts, based on the human per-
a proper subset of original feature set, in order to reduce ception of how each music is danced. The genre labeling
the preprocessing and classification steps, but maintaining was performed by two professional teachers with over 10
the final classification accuracy [3], [5]. The FS methods years of experience in teaching ballroom Latin and Brazil-
are often classified in two groups: the filter approach and ian dances.
the wrapper approach [16]. In the filter approach the fea- The experiments were carried out on stratified training,
ture selection process is carried out before the use of any validation and test datasets. In order to deal with balanced
recognition algorithm. In the wrapper approach the pattern classes, three hundred different song tracks from each genre
recognition algorithm is used as a sub-routine of the system were randomly selected.
to evaluate the generated solutions. Our primary evaluation measure is the classification ac-
We emphasize that our system employs several feature curacy. Experiments were carry out using a ten-fold cross-
vectors, according to time and space decompositions. FS validation procedure, that is, the presented results are ob-
procedure is employed in time segment vectors, allowing us tained from 10 randomly independent experiment repeti-
to compare the relative importance of the features according tions.
to their time origin. In Table 2 we present the results obtained with the ap-
Our FS procedure is based on the genetic algorithm plication of the different classifiers to the beginning music
paradigm. Individuals (chromosomes) are n-dimensional segment (first 30 seconds). Since we are evaluating the fea-
binary vectors, where n is the max feature vector size (30 ture selection procedures using the MARSYAS framework,
in our case). Fitness of the individuals are obtained from it is important to measure its performance without the use of
the classification accuracy of the corresponding classifier, any FS mechanism; this evaluation corresponds to the base-
according to the wrapper approach. line (BL) presented in the second column. Columns 3 and
4 show the results for OAA and RR space decomposition
The global feature selection procedure is as follows: approaches without feature selection; columns FS, FSOAA
1. each individual works as a binary mask for an associ- and FSRR show the corresponding results with the feature
ated feature vector; selection procedure. Results for the middle and end seg-
2. an initial assignment is randomly generated: a value 1 ments can be found in [18].
indicates that the corresponding feature is used, 0 that Analogously, Table 3 presents global results using time
it must be discarded; and space decompositions, for OAA and RR approaches,
3. a classifier is trained using the selected features; with and without feature selection. We emphasize that this
4. the generated classification structure is applied to a val- table encompasses the three time segments (beginning, mid-
idation set to determine its accuracy, which is consid- dle and end).
ered as the fitness value of this individual; Summarizing the results in Table 3, we conclude that the
5. we proceed elitism to conserve the top ranked individ- FSRR method improves classification accuracy for the clas-
uals; crossover and mutation operators are applied in sifiers J48, 3-NN and NB. Also, OAA and FSOAA methods
order to obtain the next generation. present similar results for the MLP classifier, and only for
the SVM classifier the best result is obtained without FS.
In our FS procedure we employ 50 individuals in each
As previously mentioned, we also want to analyze if dif-
generation, and the evolution process ends when it con-
ferent features have the same importance according to their
verges (no significant change in successive generations) or
when a fixed max number of generations is achieved. 1 Feature vectors available in www.ppgia.pucpr.br/˜ silla/lmd/.

42

Authorized licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.
 Table 2. Classification accuracy (%) using Table 4. Selected features in each time seg-
space decomposition for the beginning seg- ment (BME mask)
ment of the music Feature 3-NN J48 MLP NB SVM #
1 000 001 010 101 111 7
2 000 000 010 010 011 4
Classifier BL OAA RR FS FSOAA FSRR 3 000 001 010 011 000 4
J48 39.60 41.56 45.96 44.70 43.52 48.53 4 000 111 010 111 001 8
3-NN 45.83 45.83 45.83 51.19 51.73 53.36 5 000 000 110 101 100 5
MLP 53.96 52.53 55.06 52.73 53.99 54.13 6 111 101 111 111 110 13
NB 44.43 42.76 44.43 45.43 43.46 45.39 7 011 110 110 000 100 7
SVM – 23.63 57.43 – 26.16 57.13 8 001 111 110 000 111 9
9 111 111 111 111 111 15
10 110 011 111 111 111 13
Table 3. Classification accuracy (%) using 11 100 001 111 001 110 8
global time and space decomposition 12 011 010 111 011 111 11
13 111 011 111 111 111 14
14 001 010 101 000 011 6
Classifier BL OAA RR FS FSOAA FSRR 15 011 111 111 111 111 14
J48 47.33 49.63 54.06 50.10 50.03 55.46 16 111 111 111 111 111 15
3-NN 60.46 59.96 61.12 63.20 62.77 64.10 17 111 100 111 111 111 13
MLP 59.43 61.03 59.79 59.30 60.96 56.86 18 111 111 111 111 111 15
NB 46.03 43.43 47.19 47.10 44.96 49.79 19 111 010 111 111 111 13
SVM – 30.79 65.06 – 29.47 63.03 20 011 010 110 101 101 9
21 111 111 111 101 111 14
22 111 110 111 111 111 14
time origin. Table 4 shows a schematic map indicating the
23 111 111 111 100 111 13
features selected in each time segment. In this table we em-
24 011 000 111 001 011 8
ploy a binary BME mask – for (B)eginning, (M)iddle and
(E)nd time segments – where 0 indicates that the feature 25 111 011 101 111 111 13
was not selected and 1 indicated that it was selected by the 26 000 010 100 111 111 8
FS procedure in the corresponding time segment. 27 000 111 000 101 101 7
In order to evaluate the discriminative power of the fea- 28 111 111 011 111 111 14
tures, the last column in this table indicates how many times 29 000 100 000 000 101 3
the corresponding feature was selected in the experiments 30 000 011 000 111 000 5
(max 15 selections). Although this evaluation can be crit-
icized, since different features can have different impor-
tance according to the employed classifier, we argue that are selected from different time segments of the music, and
this counting gives an idea of the global feature discrimina- one-against-all and round-robin composition schemes are
tive power. For example, features 6, 9, 10, 13, 15, 16, 17, employed for space decomposition. From the partial classi-
18, 13, 21, 22, 23, 25 and 28 are important for music genre fication results originated from these views, an unique final
classification. We remember that features 1 to 6 are Beat classification label is provided. We employ a large brand of
related, 7 to 25 are related to Timbral Texture, and 26 to 30 classifiers and heuristic combination procedures in order to
are Pitch related 2 . produce the final music genre label.
An extensive set of tests were performed in order to eval-
5 Concluding Remarks uate the feature selection procedure. Our procedure is based
on the genetic algorithm paradigm, where each individual
In this paper we evaluate a feature selection procedure works as a mask that selects the set of features to be used
based on genetic algorithms in the automatic music genre in the classifier construction. The fitness of the individu-
classification task. We also use an ensemble approach ac- als are based on its classification accuracy, according to the
cording to time and space decompositions: feature vectors wrapper approach. The framework encompasses classical
genetic operations (elitism, crossover, mutation) and stop-
2 See MARSYAS [21] for a complete description of the features. ping criteria.

43

Authorized licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.
 Experiments were conducted in a new large database techniques. In Proc. of the 5th ACM SIGMM Intern. Work-
– the Latin Music Database, with more than 3,000 music shop on Multimedia Information Retrieval, pages 102–108,
pieces from 10 music genres – methodically constructed for 2003.
this research project [19], [18]. [9] J. Lee and J. Downie. Survey of music information needs,
uses, and seeking behaviours: preliminary findings. In Proc.
The results achieved with the feature selection show that
of the 5th Intern. Conf. on Music Information Retrieval,
this procedure is effective for J48, k-NN and Naı̈ve-Bayes
pages 441–446, Barcelona, Spain, 2004.
classifiers; for MLP and SVM the FS procedure does not [10] M. Li and R. Sleep. Genre classification via an lz78-based
increases classification accuracy (Tables 2 and 3); these re- string kernel. In Proc. of the 6th Intern. Conf. on Music
sults are compatible with the ones presented in [23]. Information Retrieval, pages 252–259, London, UK, 2005.
We emphasize that the use of the time/space decompo- [11] T. Li and M. Ogihara. Music genre classification with tax-
sition approach represents an interesting trade-off between onomy. In Proc. of IEEE Intern. Conf. on Acoustics, Speech
classification accuracy and computational effort; also, the and Signal Processing, pages 197–200, Philadelphia, USA,
use of a reduced set of features implies a smaller processing 2005.
time. This point is an important issue in practical applica- [12] H. Liu and L. Yu. The Handbook of Data Mining, chapter
Feature Extraction, Selection, and Construction, pages 409–
tions, where an adequate compromise between the quality
424. Lawrence Erlbaum Publishers, 2003.
of a solution and the time to obtain it must be achieved. [13] C. McKay and I. Fujinaga. Musical genre classification: Is
Another conclusion that can inferred from the experi- it worth pursuing and how can it be? In Proc. of the 7th
ments is that the features have different importance in the Intern. Conf. on Music Information Retrieval, pages 101–
classification, according to their origin music segment (Ta- 106, Victoria, CA, 2006.
ble 4). It can be seen, however, that some features are [14] A. Meng, P. Ahrendt, and J. Larsen. Improving music genre
present in almost every selection, showing they have a classification by short-time feature integration. In IEEE In-
strong discriminative power in the classification task. tern. Conf. on Acoustics, Speech, and Signal Processing,
Indeed, the origin, number and duration of the time seg- pages 497–500, Philadelphia, PA, USA, 2005.
[15] T. M. Mitchell. Machine Learning. McGraw-Hill, 1997.
ments, the use of space decomposition strategies and the
[16] L. Molina, L. Belanche, and A. Nebot. Feature selection
definition of the more discriminative features still remain algorithms: a survey and experimental evaluation. In Proc.
open questions for the automatic music genre classification of the IEEE Intern. Conf. on Data Mining, pages 306–313,
problem. Maebashi City, JP, 2002.
[17] C. Silla Jr., C. Kaestner, and A. L. Koerich. Automatic mu-
References sic genre classification using ensemble of classifiers (in por-
tuguese). In Proc. of the IEEE International Conference on
Systems, Man and Cybernetics (SMC 2007), pages 1687–
[1] J. J. Aucouturier and F. Pachet. Representing musical genre:
1692, Montreal, Canada, 2007.
A state of the art. Journal of New Music Research, 32(1):83–
[18] C. N. Silla Jr. Classifiers Combination for Automatic Music
93, 2003.
[2] J. Bergstra, N. Casagrande, D. Erhan, D. Eck, and B. Kégl. Classification (in portuguese). MSc dissertation, Graduate
Aggregate features and adaboost for music classification. Program in Applied Computer Science, Pontifical Catholic
Machine Learning, 65(2-3):473–484, 2006. University of Paraná, 2007.
[3] A. Blum and P. Langley. Selection of relevant features and [19] C. N. Silla Jr., A. L. Koerich, and C. A. A. Kaestner. The
examples in machine learning. Artificial Intelligence, 97(1- latin music database. In Proceedings of the 9th International
2):245–271, 1997. Conference on Music Information Retrieval, pages 451–456,
[4] C. H. L. Costa, J. D. ValleJr, and A. L. Koerich. Automatic 2008.
classification of audio data. In IEEE Intern. Conf. on Sys- [20] C. SillaJr., C. Kaestner, and A. L. Koerich. Time-space
tems, Man, and Cybernetics, pages 562–567, The Hague, ensemble strategies for automatic music genre classifica-
Holand, 2004. tion. In Brazilian Symposium on Artificial Intelligence (Lec-
[5] M. Dash and H. Liu. Feature selection for classification. ture Notes in Computer Science, Vol.4140), pages 339–348,
Intelligent Data Analysis, 1(1–4):131–156, 1997. 2006.
[6] J. Downie and S. Cunningham. Toward a theory of music in- [21] G. Tzanetakis and P. Cook. Musical genre classification of
formation retrieval queries: System design implications. In audio signals. IEEE Transactions on Speech and Audio Pro-
Proceedings of the 3rd Intern. Conf. on Music Information cessing, 10(5):293–302, 2002.
Retrieval, pages 299–300, 2002. [22] I. H. Witten and E. Frank. Data Mining: Practical ma-
[7] R. Fiebrink and I. Fujinaga. Feature selection pitfalls and chine learning tools and techniques. Morgan Kaufmann,
music classification. In Proc. of the 7th Intern. Conf. on San Francisco, 2nd edition, 2005.
Music Information Retrieval, pages 340–341, Victoria, CA, [23] Y. Yaslan and Z. Cataltepe. Audio music genre classifica-
2006. tion using different classifiers and feature selection methods.
[8] M. Grimaldi, P. Cunningham, and A. Kokaram. A wavelet In Proc. of the Intern. Conf. on Pattern Recognition, pages
packet representation of audio signals for music genre clas- 573–576, Hong-Kong, China, 2006.
sification using different ensemble and feature selection

44

Authorized
View publication stats licensed use limited to: UNIVERSITY OF KENT. Downloaded on May 15, 2009 at 19:00 from IEEE Xplore. Restrictions apply.