Urban Audio Classification Using Artificial
Neural Network
1st Amit Kishor Raturi
Dept. Of Computer Science And Engineering
NIT Uttarakhand
Srinagar Garhwal, India
Amitkishorraturi.cse16@nituk.ac.in
Abstract—This paper considers the use of Artificial
Neural Networks in classifying short audio clips of
environmental and urban sounds. Here we have pro-
posed the use of Artificial Neural Networks overcoming
the problem o f accuracy due to data scarcity on the
performance of the proposed Convolutional Neural Net-
work architecture. Using optimal activation functions
a fully connected 3 hidden layer network is trained
on 8,732 labelled sound clips (4 seconds each) from
ten classes: air conditioner, car horn, children playing,
dog bark, drilling, engine idling, gunshot, jackhammer,
siren, and street music. These type of sounds are
very common in daily life so classification over this
data set of environmental and urban sounds will open
the future possibilities in city management and safety
operations , especially in the smart city engineering.
Audio features which include noise, are extracted using
librosa implementation. The accuracy of the network
is evaluated on Urban8K dataset of environmental and
urban recordings.
Index Terms—environmental sound, artificial neural
network, mel-frequency cepstral coefficients
I. I NTRODUCTION
Sound recognition is a basic audio signal process-
ing problem. For instance, applications in context
aware computing [1] and surveillance [2] to noise
mitigation enabled by smart acoustic sensor networks
[3], automated route, assistive robotics, and other cell
phone based administrations, where setting mindful
handling is frequently wanted or required. We hu-
mans use both vision and audio recognition in our day
to day activities while machines are still using vision
recognition primarily as audio recognition still needs
to be done in a vast field. As till researchers have
mainly focused on similar type of audio data set, here
we have taken urban sound clips which consists of
frequent frequency changes and a lot of surrounding
noise.
At the point when utilized to comprehend unstruc-
tured situations [4], [5] (e.g., determining interior or
2nd Dhruv Verma
Dept. Of Computer Science And Engineering
NIT Uttarakhand
Srinagar Garhwal, India
Dhruv.cse16@nituk.ac.in
time_freq_audio.png
Fig. 1. Visiualization of Audio Data
exterior locations [6], [7]), machines usefulness can
be threatened with the loss of their vision recogni-
tion ability. So in such conditions audio recognition
seems very promising in increasing the optimality
of a machine. There have been ongoing interests in
nding approaches to give hearing to portable robots
[8], [9] so to improve their setting mindfulness with
sound data, Different applications incorporate those
in the area of wearables and setting mindful appli-
cations [10], [11], e.g., in the plan of a cell phone,
for example, a mobile device that can consequently
change the notication mode dependent on the learning
of client’s environment, such as changing to the
quiet mode in a theater or classroom [10] or even
give data altered to client’s area [12]. Research in
general audio environment recognition has attracted
some interest in the last few years [13],[14] but,
the action is significantly less contrasted with that
for speech or music. Programmed unstructured con-
dition characterization is still in its earliest stages.
A few regions of non-speech sound recognition that
have been concentrated to different degrees are those
relating to acknowledgment of specific occasions
utilizing sound from deliberately created movies or
TV tracks[15],[16]. Others incorporate the separation
between music instruments[17], [18], music classes
[19], and between varieties of speech, non-speech and
music [20] [22]. As till now very less systems have
proposed to model raw environmental audio, so in
this paper instead of classifying discrete voice events
we have tried to classify environmental and urban
sounds.
 II. L ITERATURE S URVEY
When contrasted with different zones in sound, for
example, discourse or music, investigate on general
unstructured sound based scene acknowledgment has
gotten little consideration. As far we know only a
few frameworks(and systems) have been proposed
to examine ecological classication with crude sound.
Sound-based circumstance examination has been con-
sidered in [23], [24] and in [11], [25], for wearables
and setting mindful applications. In view of haphaz-
ardness, high fluctuation, and different difculties in
working with ecological sounds, the acknowledgment
rates fall quickly with expanding number of classes;
agent results show acknowledgment exactness re-
stricted to around 92% for ve classes[8],77%for 11
classes [26], and roughly 60% for at least 13 classes
[23], [24]. The examination of sound conditions in
Peltonen’s postulation [24], displayed two classica-
tion plans. The rst plot depended on averaging the
band-vitality proportion as highlights and character-
izing them utilizing a K-closest neighborhood (kNN)
classier. The second uses MFCCs as highlights and a
Gaussian blend demonstrate (GMM) classier. Pelto-
nen saw the weaknesses of MFCCs for natural sounds
and proposed utilizing the band-vitality proportion
as an approach to speak to sounds happening in
various recurrence ranges. Both of these investiga-
tions included characterizing 13 unique settings or
classes. There has also been some earlier paintings
on using matching pursuit for studying audio for
classication however quite confined. The proposed
method by using Ebenezer et al. [27] demonstrated
the usage of MP for signal classication. Their frame-
work classied acoustic emissions using a modied
MP algorithm in an actual acoustic tracking machine
. The classier become based on a modied version
of the MP decomposition set of rules. For each
class, suitable gaining knowledge of signals have
been selected, the time- and frequency-transferring
of those alerts bureaucracy their dictionary. After
the MP algorithm terminates, the net contribution
of correlation coefcients from each elegance is used
because the decision statistic, wherein the only that
produces the largest cost is the chosen elegance. They
established an overall classication charge of 83.0%
for the 12-elegance classication case. In this type
of sound classification over environmental and urban
sounds, till now mainly proposed method is using
Convolutional Neural Network architecture. Karol J.
Piczaks work over environmental sound classification
with CNN is one of the example[28]. But scarcity
of labeled data for environmental sound classication
made adverse impact on final accuracy. So Justin
Salamon and Juan Pablo Bellos work Deep Convolu-
tional Neural Networks and Data Augmentation for
Environmental Sound Classication[29] shows approx-
imately the same work by removing the problem of
scarcity of labeled data for environmental sound clas-
sication using Data Augmentation and performance
of this proposed model increased signicantly, yielding
a mean accuracy of 0.79 So as earlier CNN algorithm
is mainly used in classifying these type of sound data-
set, the goal of this paper is to provide an answer
to the question that Can Artificial Neural Network
be effectively used in classifying environmental and
urban sound sources?
III. M ETHODOLOGY
A. Data Preprocessing And Feature Extraction
The raw audio signal is not suitable as direct input
to a classifier due to its extremely high dimensionality
and the fact that it would be unlikely for perceptually
similar sounds to be neighbours in vector space.Thus,
a popular approach for feature learning from audio is
to convert the signal into a time-frequency represen-
tation,a common choice being the mel-spectrogram.
We extract log-scaled mel-spectrograms with 40 com-
ponents (bands) covering the audible frequency range
(0-22050 Hz), using a window size of 23 ms (1024
samples at 44.1 kHz) and a hop size of the same
duration. We also experimented with a larger numbers
of bands (128), but this did not improve performance
and hence we stuck to the lower (and faster to
process) resolution of 40 bands.
To extract the useful features from sound data, we
have used Librosa library. Librosa provides several
methods to extract different features from the sound
clips. Following features have been extracted from
the audio clips
1) melspectrogram: Compute a Mel-scaled power
spectrogram
2) mfcc: Mel-frequency cepstral coefficients
chorma-stft: Compute a chromagram from a
waveform or power spectrogram
3) spectralcontrast Compute spectral contrast:
4) tonnetz: Computes the tonal centroid features
(tonnetz)
Librosa provides useful method to convert raw
sound clips into informative features (along with a
class label for each sound clip) that we can directly
feed into our classifier. The class label of each sound
clip is in the file name. For example, if the file name
is 108041-9-0-4.wav then the class label will be 9.
 Logistic sigmoid Input Hidden Hidden Hidden Output
1
layer layer 1 layer 2 layer 3 layer
σ(x)
(2)
h0
0.5
(1) (3)
h0 h0
x0 (2)
h1
0
−4 −2 0 2 4 (1) (3)
ŷ1
h1 h1
x
x1 (2)
h2
(a) sigmoid activation function. (1) (3)
h2 h2 ŷ2
1 Hyperbolic tangent x2 (2)
h3 ..
.
.. (1) (3)
ŷ10
tanh(x)

. h3 h3
0 x193 .. (2)
h4 ..
. .
(1)
h280 .. (3)
h250
−1
.
(2)
h300
−4 −2 0 2 4
x
Fig. 3. Proposed Architecture of classifier for urban sound
(b) Hyperbolic tangent activation function. classification

Fig. 2. Activation functions utilised in the architecture include the

sigmoid σ(x) and the hyperbolic tangent tanh(x). percent accuracy on Urban8Kdataset. There are in
total 5 layers including 3 hidden layers, input layer
B. Network Architecture and output layer in the proposed architecture. There
are 280,300,250 units in hidden layers repectively as
A typical neural network classifier consists of a shown below. The first and second layers are Tangent
number of different layers stacked together along with hyperbolic activated while third layer is sigmoid
different activation functions. activated, finally at ouput layer we have used softmax
A activation function (also known as transfer func- activation for 10 class of audios.
tion) determines the relation between input and out-
Different architechtures along with the combina-
put. Activation function normalizes the output value
tions of activation functions have been used, for
to certain range. There are number of activation
example different combinations of Rectified Linear
functions in use with neural networks. Activation
Unit(Relu), Hyperbolic Tangent function(Tanh) and
functions may be linear or non-linear. Some of the
Sigmoid activation functions have been tried, out
widely used activation functions are as follows.
a) Logistic Sigmoid Function: Sigmoid function of which the following architecture gave the best
has s-shaped graph and is most common form of ac- performance.
tivation function for defining the output of a neuron.
TABLE I
Sigmoid function is strictly increasing, continuous D ESCRIPTION O F N EURAL N ETWORK
and differential function. It exhibits a graceful bal-
ance between linear and nonlinear behavior. An ex- Layer Neurons Activation Function
Input 193 -
ample of the sigmoid function is the logistic function, Hidden 1 280 tanh
defined by, Hidden 2 300 sigmoid
1 Hidden 3 250 sigmoid
φ(x) = (1)
1 + e−x Output 10 Softmax

where x is the local output of the neuron.

b) Hyperbolic tangent sigmoid function: Hyper-
bolic tangent sigmoid activation function is symmet-
ric bipolar activation function, defined by,
2
φ(x) = −1
(1 + e−x )
We have tried the classication of urban sound with
different architectures. We have achieved around 78
1) Cost Function: A cost function or loss function
or cost function is a function that maps an event or
values of one or more variables onto a real number
intuitively representing some ”cost” associated with
(2) the event. We have used Cross-entropy loss function
described below.
Cross-entropy loss: Cross-entropy loss, or log
loss, measures the performance of a classification
cross_entropy.png
Fig. 4. Variation of Cross Entropy Loss
model whose output is a probability value between 0
and 1. Cross-entropy loss increases as the predicted
probability diverges from the actual label. So predict-
ing a probability of .012 when the actual observation
label is 1 would be bad and result in a high loss value.
A perfect model would have a log loss of 0. With
y being the target label and p being the probability
to belong to class labeled y ,cross entropy loss is
mathematicaly defined as
φ(y) = −(y log(p) + (1 − y) log(1 − p)) (3)
Its plot is as shown in fig(3).
IV. R ESULTS A ND A NALYSIS
We have splited the Urban8KDataset into 0.8 (
Training) and 0.2(Testing). Achieving an accuracy of
78% which comparatively good considering the small
dataset and no use of convolution nerual network.
We describe the performance measures that have
been used, Experiments carried out by tunning the
hyperparameters such as learning rate and batch size
and confusion matrix for proposed mdel and finally
we compare the performance of our classifier with
the convolutional nerual network implementations for
urban sound classifications in the following subsec-
tions.
1) Performance Meansures: Performance mea-
sures are a crucial part of a classification task. We
have utilised following perfromance measures for our
multi class urban audio classification.
Confusion Matrix: A confusion matrix is a table
that is often used to describe the performance of a
classification model on a set of test data for which
the true values are known. The general structure of
a confusion matrix is shown in the table(2). The
rows shows the actual class and columns show the
predicted class. Entries of the table(2) are described
TABLE II
G ENERAL S TRUCTURE O F A C ONFUSION M ATRIX
TP FP
FP FN
below
a) True Positive(TP): Number of samples that
belongs to positive class and are labeled as a member
of positive class by the classifer.
b) False Positive(FP): Number of samples that
belongs to negative class and are labeled as a member
of positive class by the classifer.
c) True Negative(TN): Number of samples that
belongs to negative class and are labeled as a member
of negative class by the classifer.
d) False Negative(FN): Number of samples that
belongs to negative class and are labeled as a member
of positive class by the classifer.
Based on the above parameters follwing per-
formance measures are defined for a classification
model. Accuracy defines the percentage of correct
predictions on test data.
TP + TN
Accuracy = (4)
TP + TN + FP + FN
Precision is defined as the ration of total number
of correctly classified positive examples with the
total number of predicted positive examples. High
Precision indicates an example labeled as positive is
indeed positive (small number of FP).
TP
P recision = (5)
TP + FP
Recall can be defined as the ratio of the total number
of correctly classified positive examples divide to
the total number of positive examples. High Recall
indicates the class is correctly recognized (small
number of FN).
TP
Recall = (6)
TP + FN
Since we have two measures (Precision and Recall)
it helps to have a measurement that represents both
of them. We calculate an F-measure which uses
Harmonic Mean in place of Arithmetic Mean as it
punishes the extreme values more.
2 ∗ P recsion ∗ Recall
F − M easure = (7)
P recision + Recall
2) Experimental Results: Urban sound classifica-
tion with fully connected neural network achieves
competitve accuracy as compared to CNN based
urban sound classifiers. We are able to get 78%
of accuracy with the proposed architecture. Perfor-
mance measures for our classifier is shown in ta-
ble(3). Confusion matrix for our experiments with
Urban8KDataset are shown in the fig(4). The cost
as a function of number of epochs is shown in the
fig(5).

Fig. 5. Confusion Matrix for urban sound classifier
iteration_vs_cost_function.png
Fig. 6. Cost Function vs Iterations
V. C ONCLUSION
The goal of this paper was to evaluate whether
fully connected neural networks can be successfully
applied to environmental sound classification tasks,
especially considering the limited nature of datasets
available in this field. It seems that they are in-
deed a viable solution to this problem. Conducted
experiments show that a fully connected neural net-
work model compete with other comman approaches
like applying convolution nerual network where the
scarcity of data can be a major issue. We are able
to produce an accuracy of 78%. Although, taking
into consideration much longer training times, the
result is far from ground- breaking, it shows that fully
connected neural networks can be effectively applied
in environmental sound classification tasks even with
limited datasets. What is more, it is quite likely that
a considerable increase in the size of the available
dataset would vastly improve the performance of
trained models, as the gap to human accuracy is
still profound. One of the possible questions open
for future inquiry is whether fully connected neural
networks could outperform the convolution nerual
network approaches.
ACKNOWLEDGMENT
We are highy greatful to Dr. Maroti Deshmukh ,
Assistant Professor ( Dept. Of Computer Science &
Engineering , NIT Uttarakhand) for their guidence
during the course of this research project.
TABLE III
P ERFORMANCE M EASURES F OR PROPOSED CLASSIFIER
Accuracy = 0.783 Precision = 0.731
Recall = 0.723 F-measure = 0.78
R EFERENCES
[1] S. Chu, S. Narayanan, and C.-C. Kuo, Environmental sound
recognition with time-frequency audio features, IEEE Trans.
on Audio, Speech, and Language Processing, vol. 17, no. 6,
pp. 11421158, Aug. 2009.
[2] R. Radhakrishnan, A. Divakaran, and P. Smaragdis, Audio
analysis for surveillance applications, in IEEE Worksh. on
Apps. of Signal Processing to Audio and Acoustics (WAS-
PAA05), New Paltz, NY, USA, Oct. 2005, pp. 158161.
[3] C. Mydlarz, J. Salamon, and J. P. Bello, The implementation
of low cost urban acoustic monitoring devices, Applied
Acoustics, vol.InPress, 2016.
[4] J.Pineau,M.Montemerlo,M.Pollack,N.Roy,andS.Thrun,Towards
robotic assistants in nursing homes: Challenges and results,
Special Iss. Socially Interactive Robots, Robot., Autonomous
Syst., vol. 42, no. 34, pp. 271281, 2003.
[5] S. Thrun, M. Bennewitz, W. Burgard, A. B. Cremers, F.
Dellaert, D. Fox, D. Haehnel, C. Rosenberg, N. Roy, J.
Schulte, and D. Schulz, Minerva: A second generation mobile
tour-guide robot, in Proc. ICRA, 1999.
[6] H.A.Yanco,A robotic wheel chair system: Indoor navigation
and user interface, in Lecture Notes in Articial Intelligence:
Assistive Technology and Articial Intelligence. NewYork:
Springer-Verlag,1998, pp. 256268.
[7] A. Fod, A. Howard, and M. J. Mataric, Laser-based people
tracking, in Proc. ICRA, 2002.
[8] S. Chu, S. Narayanan, C.-C. J. Kuo, and M. J. Mataric ,
Where am I? Scene recognition for mobile robots using audio
features, in Proc. ICME, 2006.
[9] J. Huang, Spatial auditory processing for a hearing robot, in
Proc. ICME, 2002.
[10] A. Waibel,H. Steusloff,andR. Stiefelha-
gen,ChilComputersinthe humaninteraction-
loop,inProc.WIAMIS,2004,andtheCHILProject Consortium.
[11] D. P. W. Ellis and K. Lee, Minimal-impact audio-based
personal archives, in Proc. CARPE, 2004.
[12] J. Mantyjarvi, P. Huuskonen, and J. Himberg, Collaborative
context determination to support mobile terminal applica-
tions, IEEE Trans. Wireless Communications, vol. 9, no. 5,
pp. 3945, Oct. 2002.
[13] D. P. W. Ellis, Prediction-driven computational auditory
scene analysis, Ph.D. dissertation, Dept. of Elect. Eng. and
Comput. Sci., Mass. Inst. of Technol., Cambridge, MA, Jun.
1996.
[14] J.-J. Aucouturier, B. Defreville, and F. Pachet, The bag-of-
frames approach to audio pattern recognition: A sufcient
model for urban soundscapes but not for polyphonicmusic, J.
Acoust.Soc. Amer., vol. 122, no. 2, pp. 881891, Aug. 2007.
[15] R. Cai, L. Lu, A. Hanjalic, H. Zhang, and L.-H. Cai, A exible
framework for key audio effects detection and auditory con-
text inference, IEEE Trans. Audio, Speech, Lang. Process.,
vol. 14, no. 3, pp. 10261039, May 2006.
[16] R. Radhakrishnan, A. Divakaran, and P. Smaragdis, Audio
analysis for surveillance applications, in Proc. IEEE Work-
shop Applicat. Signal Process. Audio Acoust., 2005, pp.
158161.
[17] P. Cano, M. Koppenberger, S. Le Groux, J. Ricard, N. Wack,
and P. Herrera, Nearest-neighbor generic sound classication
with a wordnet-based taxonomy, in Proc. 116th AES Conv.,
2004.
[18] P. Herrera, A. Yeterian, and F. Gouyon, Automatic classi-
cation of drum sounds: A comparison of feature selection
methods and classication techniques, in Proc. ICMAI, 2002.
[19] G. Tzanetakis and P. Cook, Musical genre classication of
audio signals, IEEE Trans. Speech Audio Process., vol. 10,
no. 5, pp. 293302, Jul. 2002.
[20] M. J. Carey, E. S. Parris, and H. Lloyd-Thomas, A compar-
ison of features for speech, music discrimination, in Proc.
ICASSP, 1999.
[21] K. El-Maleh, M. Klein, G. Petrucci, and P. Kabal,
Speech/music discrimination for multimedia applications, in
Proc. ICASSP, 2000, pp. 149152.
[22] T.ZhangandC.-C.J.Kuo, Audio content analysis for online
audio visual data segmentation and classication, IEEE Trans.
Speech Audio Process., vol. 9, no. 4, pp. 441457, May 2001.
[23] A. Eronen, V. Peltonen, J. Tuomi, A. Klapuri, S. Fagerlund,
T. Sorsa, G. Lorho, and J. Huopaniemi, Audio-based context
recognition, IEEE Trans. Audio, Speech, Lang. Process., vol.
14, no. 1, pp. 321329, Jan. 2006.
[24] V. Peltonen, Computational auditory scene recognition, M.S.
thesis, Tampere Univ. of Technol., Tampere, Finland, 2001.
[25] B. Clarkson, N. Sawhney, and A. Pentland, Auditory con-
text awareness via wearable computing, in Proc. Workshop
Perceptual User Interfaces, 1998.
[26] R. G. Malkin and A. Waibel, Classifying user environment
for mobile applications using linear auto encoding of ambient
audio,in Proc. ICASSP, 2005
[27] S. P. Ebenezer, A. Papandreou-Suppappola, and S. B. Sup-
pappola, Classication of acoustic emissions using modied
matching pursuit, EURASIP J. Appl. Signal Process, pp.
347357, 2004.
[28] Karol J. Piczak, Environmental sound classication with
Convolutional Neural Networks, IEEE INTERNATIONAL
WORKSHOP ON MACHINE LEARNING FOR SIGNAL
PROCESSING, SEPT. 1720, 2015.
[29] Justin Salamon and Juan Pablo Bello, Deep Convolutional
Neural Networks and Data Augmentation for Environmen-
tal Sound, IEEE SIGNAL PROCESSING LETTERS, AC-
CEPTED NOVEMBER 2016.