Automatic Acoustic Scene Classification


Automatic Acoustic Scene Classification

Gonçalo Marquesa, Thibault Langloisb
aISEL, Electronic Telecommunications and Computer Department, Lisbon, Portugal

bFCUL, Informatics Department, Lisbon, Portugal
gmarques@deetc.isel.pt tl@di.fc.ul.pt

Abstract— This paper presents a baseline system for automatic
acoustic scene classification based on the audio signals alone. The
proposed method is derived from classic, content-based, music
classification approaches, and consists in a feature extraction
phase followed by two dimensionality reduction steps (principal
component analysis and linear discriminant analysis) and a
classification phase done using a k nearest-neighbors algorithm.
This paper also reports on how our system performed in the
context of the DCASE 2016 challenge, for the acoustic scene
classification task. Our method was ranked fifteenth amongst
forty nine contest entries, and although it is below the top
performing algorithms, in our perspective it is still interesting
to see a low-complexity system such as ours obtain fairly good
performances.

Keywords: Machine Learning, Signal Processing, Music
Information Retrieval, Bag of Frames.

I. INTRODUCTION

Automatic identification of sound sources in an urban envi-
ronment has a huge potential in several applications related to
the current panorama of intelligent cities. These applications
include monitoring systems able to recognize activities, sound
environments, and create city sound maps to provide to the
general public information about environmental noise, or other
acoustic factors. However, a lot of research is still needed to
reliably detect and recognize sound events and scenes in real-
istic environments where multiple sources, often distorted, are
simultaneously present. This work focuses on one particular
aspect of urban sound analysis: acoustic scene classification.

The system we propose is a classical classification sys-
tem in the sense that it uses typical machine-learning data
transformation and classification algorithms in the decision
making process. First, each audio excerpt is converted into a
single feature vector which is the representation of choice for
standard machine-learning methods. Then, the whole dataset
is transformed via principal component analysis (PCA) [10],
an unsupervised dimensionality reduction technique, followed
by a linear discriminant analysis (LDA) projection [9]. LDA
is a supervised process, and the projection tries to maximize
the ratio between intra and inter class scatter, but it is not
a classification method since no decision is involved. For
classification, a k-nearest neighbors (k-NN) algorithm was
used [7]. The experimental configuration used in our tests
is common in many audio classification works (or at least
parts of it – see for instance [6], [11], [14], [16], [18]) and
therefore it does not bring any original contribution in terms
of the algorithmic setup. In fact, our system falls under the
standard “bag of frames” classifiers commonly used in music

information retrieval applications (see [3], [4], [12], [17] and
references therein). Our main objective was not to bring forth
a new audio classification or feature extraction method, but
rather see how a simple, non parametric algorithm performed
in the acoustic scene classification challenge. We used the
same data partitioning and cross-validation setup provided
with the database and our results are a bit better than the
ones reported in [13]. The method in [13] was the baseline
system provided with the challenge and ranked twenty eighth
while our system ranked fifteenth (amongst a total of forty nine
entries). However, the experiments we conducted also revealed
some unexpected variations in terms of accuracy when the
whole dataset or just part of it was used to estimate the PCA
and LDA projections. This is an indication that there may
be differences between feature class-dependent distributions
among folds. The structure of the remainder of this paper
is as follows: Section II describes the data and the feature
extraction process used in our experiments, Section III de-
scribes our approach to acoustic scene classification, followed
by Section IV where we present our results. In this section, we
also report on the challenge results and how we faired against
other contestants. Section V concludes this paper.

II. DATA AND FEATURE EXTRACTION

The dataset used in this work was created in the context
of the DCASE2016 challenge [1] for the acoustic scene
classification task. The dataset contains 1170, 30-seconds
audio excerpts from the following acoustic scenes: Beach, Bus,
Café/Restaurant, Car, City Center, Forest Path, Grocery Store,
Home, Library, Metro Station, Office, Urban Park, Residential
Area, Train, and Tram. The dataset is divided into four folds
for cross-validation testing. We used the same data partition
in our experiments and our results are averaged over the four
test folds.

The features used are the all-purpose Mel frequency cepstral
coefficients (MFCCs), a very popular representation in speech
recognition (see for instance [15]), and also widely used in
content-based music information applications. The audio was
decomposed into 23 ms segments (1024 samples at 44.1 kHz)
with 50% overlap, and we used 100 Mel bands to extract 23
MFCCs plus the zero order MFCC and the frame’s log-energy,
plus the delta and acceleration coefficients. This means that the
audio is converted into a sequence of 25×3=75 dimensional
vectors. We applied the VoiceBox software [2] to extract the
features. In order to convert each audio excerpt into a single
feature vector, the sequence of MFCC features is summarized

i-ETC: ISEL Academic Journal of Electronics, 
Telecommunications and Computers 
CETC2016 Issue, Vol. 3, n. 1 (2017) ID-2 
 

http://journals.isel.pt 



using the median and logarithmic standard deviation. The
median was used instead of the mean since this statistic is
more robust to outliers. The log-standard deviation is given
by 20 log10(σi) where σi is the standard deviation of feature
i (with i = 1,. . . ,75). The reason to use the log-standard
deviation instead of plain standard deviation was to convert
these feature values to an order of magnitude comparable to
the median feature values - otherwise the standard deviation
values would be a few orders of magnitude lower, and during
the PCA pre-processing step, these dimensions would be
discarded as noise since they would not contribute in any
significant way to the overall data variance. The statistics
return two 75-dimensional vectors which are concatenated, so
each audio excerpt is represented by a 150-dimensional feature
vector.

III. METHOD

The proposed classification approach is divided into three
main blocks: feature pre-processing via principal component
analysis, feature transformation by linear discriminant analysis
and finally a classification step performed by a k-nearest
neighbor classifier.

Principal Component Analysis: PCA is a standard dimen-
sionality reduction technique, where the data is decorrelated by
projecting it into orthogonal directions of maximum variance.
These directions, the principal components, are obtained using
a eigen-decomposition of the data covariance matrix, and in
our experiments we kept enough components to explain 99.9%
of the total data variance. The PCA-transformed data was also
whitened - each data dimension was scaled in order to have
unit variance.

Linear Discriminant Analysis: LDA is commonly used
as a pre-processing step for pattern classification. It is also a
dimensionality reduction technique since the data is projected
into c −1 dimensional space where c is the total number of
classes (c=15 for this challenge).

LDA is a supervised learning method, and therefore the
projection should be calculated with the training set only,
otherwise we are indirectly including information about the
class labels in the test set. Estimating the LDA projection with
the whole dataset can result in overly optimistic performance
values, specially when there is a relatively large number of
classes and a relatively low number of examples, as in the
case of this challenge dataset. We tested the performance
of our system using the whole dataset to estimate the LDA
projection in order to assess the increase in performance
compared to the “correct” evaluation procedure. The results
showed a significant increase in performance, which in our
perspective, is somewhat surprising. These are reported in
Section IV-B, along with a discussion on possible causes of
such a performance discrepancy.

k-Nearest Neighbors: k-NN is an instance-based learning,
where class membership of a pattern is assigned based on a
majority vote of its neighbors. k-NN is possibly one of the
simplest classification methods, and therefore it is well suited
for a baseline system. We tested two distance metrics with the

k-NN algorithm, the cosine and the Euclidean distance, and
opted for the Euclidean distance because it yielded slightly
better accuracy results. We also ran the algorithm with differ-
ent number of neighbors (from 5 to 31 - using an increment
of two) and chose empirically k = 9. The results reported in
Section IV are obtained using the Euclidean distance metric,
and k=9.

IV. EXPERIMENTAL RESULTS

This section is divided into three subsections. In the first, we
present the results obtained with our method. The experimental
setup is described, the system performance is measured in
terms of accuracy, either with mean or class specific values.
In the second, we present the performances obtained when the
whole dataset is used to estimate the LDA projection. This is
not the correct procedure to estimate our system performance.
The intent is to have an idea of by how much the performance
values are inflated. In the third part, we give a brief overview
of the results on the DCASE-2016 acoustic scene classification
challenge.

A. System Performance

The results presented in this section were obtained using
the following experimental setup. The PCA and the LDA
projections were estimated using only the training set. In
our tests, we used 4-fold cross validation and the same data
partitioning provided with the dataset. This means that a
total of four LDA projections where estimated with three
training folds. The presented result pertain to the tests folds
only. The obtained average accuracy was 77.6%. In Table I
reports the (average) accuracies per class. These context-
wise performances vary from 52.6% (Train class) to 93.6%
(City Center and Metro Station classes). Table II shows the
accuracies per fold.

TABLE I
Accuracy per class. Accuracy values obtained with the mean of all
four test folds.

1. Beach 76.9%
2. Bus 66.7%
3. Café/Restaurant 79.5%
4. Car 84.6%
5. City Center 93.6%
6. Forest Path 87.2%
7. Grocery Store 82.1%
8. Home 64.1%
9. Library 87.2%
10. Metro Station 93.6%
11. Office 92.3%
12. Urban Park 60.3%
13. Residential Area 65.4%
14. Train 52.6%
15. Tram 87.2%

Figure 1 shows the confusion matrix (obtained summing the
four confusion matrices in each test fold). Each line refers to
the examples of a single class; the class order is the same as the
one in Table I. The columns refer to the classification results.

G.Marques et al. | i-ETC -  CETC2016 Issue, Vol. 3, n. 1 (2017) ID-2
 

i-ETC: ISEL Academic Journal of Electronics, Telecommunications and Computers 



Fig. 1. Confusion matrix - the rows represent the true classes, the
columns represent the classification results. The class order is the
same as the one given in Table I: in the first row are the samples
from the class Beach, in the second from class Bus, and so on. This
matrix was obtained by the sum of the four confusion matrices -
one per test fold. A perfect classification would result in this matrix
having the value 78 on the main diagonal (number of examples per
class) and zeros for the rest of the entries.

60 0 0 0 1 2 0 0 1 0 1 9 3 0 1
0 52 6 2 0 0 1 0 5 0 0 0 0 11 1
0 0 62 0 0 2 8 0 2 4 0 0 0 0 0
0 3 0 66 0 0 0 0 1 0 0 0 0 7 1
1 0 0 0 73 0 0 0 1 1 0 0 2 0 0
1 0 0 0 0 68 0 0 0 0 2 7 0 0 0
0 0 4 0 0 0 64 0 1 9 0 0 0 0 0
2 2 7 0 0 2 0 50 8 0 6 0 0 0 1
0 0 9 0 0 0 1 0 68 0 0 0 0 0 0
0 0 0 0 0 0 2 0 3 73 0 0 0 0 0
0 0 0 0 0 1 0 5 0 0 72 0 0 0 0
2 1 0 0 1 13 3 0 6 0 0 47 5 0 0
2 0 2 0 1 3 0 0 0 0 1 17 51 0 1
0 11 8 0 0 2 4 0 0 5 1 0 0 41 6
0 1 0 0 0 0 7 0 0 0 1 0 0 1 68

For example, for the class Beach, 60 audio excerpts were
correctly classified, 9 were classified as the class Urban Park,
and nine others were also misclassified. The results also reveal
some particular error correlations among certain classes. For
instance many Residential Area samples were misclassified
as Urban Park (a total of 17 errors). Five Urban Park pieces
were attributed to Residential Area, however the excerpts of
this class have a higher tendency to get confused with another
class: Forest Path (a total of 13 errors). These mislabeling
seems understandable since these acoustic scenes share some
resemblances. Another example of classes that have similar
acoustic characteristics and a high number of errors between
them are Bus and Train. Other relations that seem to make
some sense could be found such as the case of Beach and
Urban Park, or Home and Library, but further tests would be
needed to determine if a real correlation exists.

TABLE II
Accuracy per fold. The mean accuracy is 77.4%. These results were
obtained using only the training set to estimate the PCA projection.

Fold 1 Fold 2 Fold 3 Fold 4
Accuracy 79.3% 71.7% 82.6% 76.0%

B. LDA estimation with the whole dataset

In this section, we discuss the results obtained when we
used the whole dataset to estimate LDA projection. This is
not the correct testing methodology, since when doing this,
we are implicitly including test label information in the model
training process. The intent here is just to determine by how
much the performances are over evaluated when using this
incorrect experimental procedure.

We performed some tests in order to have an idea of how
the classification performance is affected by using just part
or the whole dataset. The results (see Table III) show that
there is no significant decrease in accuracy (less than 1%)
when the PCA projection is estimated with only the training
set. Nevertheless, when we applied the same methodology to
estimate the LDA projection, the results were significantly
affected. Table IV shows the accuracies per test fold. The
mean accuracy is 90.8% which is 12.6% points higher than
our baseline system. Since LDA is a supervised technique, an
increase in performance is expected when the entire dataset is
used, but in our perspective such a high bias was unforeseen.
This is also an indication that there is some variability of class-
dependent feature distributions among folds. The partition
process used for this dataset may be the cause, since it was
based on recording location [13]. This division was done in
order to avoid overestimating systems performances, since in
this way, segments from a single recording are assigned to
only one fold. We believe that the variability between folds is
also due to the relatively low number of examples per class,
and increasing the number of examples in the database will
reduce this variation.

Figure 2 shows the confusion matrix. There is some similar-
ities to the error patterns found in Section IV-A. For instance,
the Residential Area samples are still misclassified as Urban
Park, Urban Park as Forest Path, and Home as Library. Other
errors though, like the confusion between Bus and Train
classes, have almost vanished.

TABLE III
Accuracy per fold. The mean accuracy is 78.2%. These results were
obtained using the whole dataset to estimate the PCA projection.
The whole dataset was used to measure the increase in performance
compared to using only the training set for the PCA estimations
(results in Table II).

Fold 1 Fold 2 Fold 3 Fold 4
Accuracy 79.0% 72.1% 82.9% 78.8%

TABLE IV

Accuracy per fold. The mean accuracy is 90.8%. These results were
obtained using (inappropriately) the whole dataset to estimate the
LDA linear projection.

Fold 1 Fold 2 Fold 3 Fold 4
Accuracy 95.9% 85.9% 92.3% 89.0%

C. DCASE 2016 Challenge Results

The DCASE 2016 challenge, provided two datasets for
acoustic scene classification. The first one, described in Sec-
tion II, was created for developing purposes and the results
reported in the previous sections were based on this dataset.
A second dataset consisting of 390 audio excerpts belonging to
one of 15 classes was also provided without the ground truth

G.Marques et al. | i-ETC -  CETC2016 Issue, Vol. 3, n. 1 (2017) ID-2
 

i-ETC: ISEL Academic Journal of Electronics, Telecommunications and Computers 



Fig. 2. Confusion matrix obtained by the sum of the four confusion
matrices - one per test fold. The results were obtained using (inap-
propriately) the whole dataset to estimate the LDA linear projection.

67 0 0 0 1 2 1 0 1 0 0 2 3 0 1
0 72 3 1 0 0 0 0 1 0 0 0 0 1 0
0 0 77 0 0 0 1 0 0 0 0 0 0 0 0
0 1 0 77 0 0 0 0 0 0 0 0 0 0 0
1 0 0 0 76 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 75 0 0 0 0 0 3 0 0 0
0 0 1 0 0 0 73 0 0 4 0 0 0 0 0
1 1 2 0 0 0 1 61 8 0 3 0 0 0 1
0 0 2 0 0 0 0 0 76 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 77 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 78 0 0 0 0
2 0 0 0 0 7 2 0 4 0 0 62 1 0 0
1 0 0 0 0 2 0 0 2 0 0 14 58 0 1
0 1 3 0 0 0 3 0 1 2 1 0 0 60 7
0 0 0 0 0 0 3 0 0 0 1 0 0 1 73

for evaluation purposes. Our submission obtained an accuracy
of 83.1%, and ranked 15thplace among 49 submissions (see
Table V for the top 20 results). The first place submission [8]
achieved a 89.7% accuracy, with a fairly complex approach.
Their system was based on a combination of different scores
obtained with distinct features such as spectrograms or I-
vectors [5] (using the audio stereo information), and also
distinct classifiers such as Deep Convolutional Neural Net-
works or creating an Universal Background Model for the I-
vectors using a mixture of Gaussians. In the end, each audio
piece was represented by 16 different scores, which were
fused together in order to provide the class estimate. The
majority of other submissions also used some sort of neural
network (NN) for classification (recurrent-NN, convolutional-
NN, deep-NN), and various sets of features. The second
method of choice was support vector machines (SVM), while
the third was a combination of various classification algorithms
through fusion or ensemble methods. In terms of algorithmic
complexity, our method is by far the simplest, and in that
sense, we believe that is interesting that such low-complexity
system can obtain fairly reasonable performances.

V. CONCLUSION

In this work, we presented a baseline for acoustic scene
classification system composed of two dimensionality reduc-
tion transformation (PCA and LDA) followed by a k-NN
classification algorithm.We trained and tested our method on
the DCASE 2016 acoustic scene classification dataset, and
submitted it to the challenge provided by the organization.
Our approach was not the top ranked one: 6.6% points below
the 1st place. Nevertheless, the performance obtained (83.1%
accuracy) is still a relatively high value, specially taking
into consideration the simplicity of the method. These results
also highlight the benefits of pre-processing the data with
dimensionality reduction techniques.

TABLE V

Top 20 ranking positions in terms of accuracy scores for the acoustic
scene classification task of the DCASE 2016 challenge (for de-
tails see their web page: http://www.cs.tut.fi/sgn/arg/dcase2016/
task-acoustic-scene-classification.)

Rank Acc. Author Classifier
1 89.7% E-Zadeh et al. fusion
2 88.7% E-Zadeh et al. I-vector
3 87.7% Bisot et al. NMF
4 87.2% Park et al. fusion
5 86.4% E-Zadeh et al. I-vector
5 86.4% Marchi et al. fusion
6 86.2% Valenti et al. CNN
7 85.9% Elizalde et al. SVM
8 85.6% Takahashi et al. DNN-GMM
9 85.4% Kim & Lee CNN-ensemble

10 84.6% Han & Lee CNN
11 84.1% Bae et al. CNN-RNN
11 84.1% Wei et al. ensemble
12 83.8% Liu et al. fusion
13 83.6% Liu et al. fusion
14 83.3% E-Zadeh et al. CNN
14 83.3% Pham et al. CNN
14 83.3% Lidy & Schindler CNN
15 83.1% Bao et al. fusion
15 83.1% Marques & Langlois k-NN
16 82.3% Mun et al. DNN
16 82.3% Wei et al. ensemble
17 82.1% Yun et al. GMM
17 82.1% Rakotomamonjy SVM
18 81.8% Lidy & Schindler CNN
19 81.3% Ghodasara et al. SVM
20 81.0% Kong et al. DNN
20 81.0% Nogueira SVM

In our tests, we also found a large variation in accuracies
when the LDA transformation was estimated using the whole
dataset versus using only the training set. This is an indication
that there is some variability of feature distribution among
folds in this particular dataset.

ACKNOWLEDGEMENT

Gonçalo Marques thanks FI-Sonic1 for supporting his re-
search.

REFERENCES

[1] DCASE 2016 - Detection and Classification of Acoustic
Scenes and Events - http://www.cs.tut.fi/sgn/arg/
dcase2016/.

[2] VOICEBOX: Speech Processing Toolbox for MATLAB (2005)
by Mike Brookes - http://www.ee.ic.ac.uk/hp/
staff/dmb/voicebox/voicebox.html.

[3] J.-J. Aucouturier and F. Pachet. Improving timbre similarity:
How high is the sky? Journal of Negative Results in Speech
and Audio Sciences, 1(1), 2004.

[4] M. Casey, R. Veltkamp, M. Goto, M. Leman, C. Rhodes, and
M. Slaney. Content-based music information retrieval: Current
directions and future challenges. Proceedings of the IEEE,
96(4):668–696, April 2008.

[5] N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet.
Front-end factor analysis for speaker verification. Trans. Audio,
Speech and Lang. Proc., 19(4):788–798, May 2011.

1http://www.fi-sonic.com/

G.Marques et al. | i-ETC -  CETC2016 Issue, Vol. 3, n. 1 (2017) ID-2
 

i-ETC: ISEL Academic Journal of Electronics, Telecommunications and Computers 



[6] S. Dieleman and B. Schrauwen. Multiscale approaches to
music audio feature learning. In 14th International Society for
Music Information Retrieval Conference (ISMIR), pages 116–
121, 2013.

[7] R. Duda, P. Hart, and D. Stork. Pattern Classification (2nd
Edition). Wiley-Interscience, 2000.

[8] H. Eghbal-Zadeh, B. Lehner, M. Dorfer, and G. Widmer. A
hybrid approach using binaural i-vectors and deep convolutional
neural networks. Technical report, DCASE2016 Challenge,
September 2016.

[9] R. Fisher. The use of multiple measurements in taxonomic
problems. Annals of Eugenics, 7(2):179–188, 1936.

[10] H. Hotelling. Analysis of a complex of statistical variables
with principal components. Journal of Educational Psychology,
24:417–441, 1933.

[11] C.-H. Lee, J.-L. Shih, K.-M Yu, and H.-S. Lin. Automatic
music genre classification based on modulation spectral anal-
ysis of spectral and cepstral features. IEEE Transactions on
Multimedia, 11(4):670–682, 2009.

[12] G. Marques, T. Langlois, F. Gouyon, M. Lopes, and M. Sordo.
Short-term feature space and music genre classification. Journal
of New Music Research, 40(2):127–137, 2011.

[13] A. Mesaros, T. Heittola, and T. Virtanen. TUT database for
acoustic scene classification and sound event detection. In 24th
European Signal Processing, Budapest, Hungary, 2016.

[14] S. R. Ness, A. Theocharis, G. Tzanetakis, and L. G. Martins.
Improving automatic music tag annotation using stacked gen-
eralization of probabilistic SVM outputs. In Proc. of the 17th
ACM Int. Conf. on Multimedia (ACM-MM’9, New York, U.S.A.,
2009. ACM.

[15] L. Rabiner and B.-H. Juang. Fundamentals of Speech Recogni-
tion. Prentice Hall, 1993.

[16] J. Salamon and J. P. Bello. Unsupervised feature learning for
urban sound classification. In IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), pages
171–175. IEEE, 2015.

[17] B. Sturm. A survey of evaluation in music genre recognition.
Proc. Adaptive Multimedia Retrieval, Denmark, 2012.

[18] G. Wu, J. Zhu, and H. Xu. A hybrid visual feature extraction
method for audio-visual speech recognition. In 16th IEEE
International Conference on Image Processing (ICIP), pages
1829–1832, 2009.

G.Marques et al. | i-ETC -  CETC2016 Issue, Vol. 3, n. 1 (2017) ID-2
 

i-ETC: ISEL Academic Journal of Electronics, Telecommunications and Computers