Applied Acoustics 202 (2023) 109168

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

CNN hyper-parameter optimization for environmental sound

classification
_
Özkan Inik
Department of Computer Engineering, Tokat Gaziosmanpasa University, Tokat, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: Environmental sounds are being used widely in our lives. It is especially used in tasks such as managing
Received 10 May 2022 smart cities, location determination, surveillance systems, machine hearing, and environmental monitor-
Received in revised form 23 November 2022 ing. The main method for this, environmental sound classification (ESC), has been increasingly studied in
Accepted 7 December 2022
recent years. However, the classification of these sounds is more difficult than other sounds because there
Available online 15 December 2022
are too many parameters that generate noise. The study tried to find the convolutional neural network
(CNN) model that gave the highest accuracy for ESC tasks with the optimization of hyper-parameters.
Keywords:
For this purpose, the Particle Swarm Optimization (PSO) algorithm was rearranged to represent the
Environmental sound classification (ESC)
CNN
CNN architecture. Thus, the hyper-parameters in CNN are represented exactly without any transforma-
Particle swarm optimization (PSO) tion during optimization. Studies were carried out on the ESC-10, ESC-50, and Urbansound8k data sets,
Hyper-parameter optimization which are state-of-art for ESC tasks. Some data augmentation techniques have been used for data sets
Urbansound8k in the training of CNN models. The CNN models, which were obtained with PSO, achieved success rates
ESC-50 of 98.64 % for ESC-10, 93.71 % for ESC-50, and 98.45 % for Urbansound8k, respectively. These results are
the best accuracy values obtained with the pure CNN model when compared with previous studies. As a
result, it has been made possible to automatically design CNN models for the classification of urban
sounds, giving high classification accuracy. Thus, researchers who do not know much about CNN design
can use this method in their desired datasets without the need for expert knowledge.
Ó 2022 Elsevier Ltd. All rights reserved.

1. Introduction
Sound data contain more semantic information than visual data
[1]. In particular, sound data becomes more important to obtain
information about an environment. To realize some applications in
daily life, it is necessary to use environmental sounds, unlike speech
and music sounds. For this reason, studies on the classification of
urban sounds have intensified in recent years. Environmental sound
classification (ESC), is known as one of the most important issues of
the non-speech voice classification task [2]. ESC is of critical impor-
tance in many problems such as; noise pollution analysis[3,4],
surveillance systems [5–7], context-aware applications [1,8–13],
machine hearing [14–17], environmental monitoring [18], crime
alert systems[19], soundscape assessment [20,21], and smart cities
[22,23]. Different data sets have been created for the ESC task.
ESC-10, ESC-50 [24], and Urbansound8k(US8K) [25] datasets are
used extensively. Different statistical and machine learning meth-
ods have been used for ESC in the literature [1,26–33].
The success rates of these methods are relatively low compared
to deep-learning-based studies. Deep learning[34] achieved a high
E-mail address: ozkan.inik@gop.edu.tr
success rate in the ImageNet [35] competition in 2012. Due to this
success, deep learning models have been used frequently in recent
years for ESC tasks [36–50]. These studies are described in the next
section.
1.1. Related work
In the study by Piczak[36], the CNN model was used to classify-
three different data sets. The proposed CNN model consists of two
convolution layers, a pooling layer, and two fully connected layers.
In the results obtained, it has been shown that the CNN model per-
forms better than other existing methods. In the study by Salamon
et al.[37], CNN was used for the classification of ESC. Data augmen-
tation was also applied in CNN training. Because of the training, it
has been shown that the CNN model performs better with data
augmentation. The CNN model was proposed for the detection of
acoustic events by Takahashi et al. [38]. The proposed CNN model
is inspired by VggNet [51]. In addition, a new method was pro-
posed for increasing the data given in training processes. In the
training phase of deep learning models, a different strategy was
developed by Tokozume et al.[39] to feed the data to the model.
In this method called Between-Class learning, sounds between

https://doi.org/10.1016/j.apacoust.2022.109168
0003-682X/Ó 2022 Elsevier Ltd. All rights reserved.
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

classes are produced by mixing two sounds belonging to different

classes at random rates. In the study, it was stated that the cross-
class learning method performed well in voice recognition net-
works and data augmentation. In the study, a network was also
defined for the classification of ESC and trained with the proposed
method. It has been stated that the results obtained are lower than
the human voice recognition error value.
Classification was performed on ESC-10, ESC-50 and US8K data-
sets by Boddapatia et al. [40] using AlexNet and GoogLeNet mod-
els. Audio signals in datasets were converted into images using
spectrogram, MFCC and CRP methods. Then, the classification pro-
cess was carried out using deep learning models on the images
obtained. An original stacked CNN model was proposed in the
study by Li et al. [41]. Multiple convolution layers are used with
reduced filter numbers in the proposed network. In the method,
two different CNN networks are trained using raw audio signal
data. These two models were then combined using Dempster–Sha-
fer (DS) evidence theory, and a new model for CNN named DS-CNN
was proposed. It is stated that the proposed model outperforms
other CNN-based models in ESC-10, ESC-50 and US8K datasets.
In the study by Su et al. [42], it was suggested to combine two
different features of ESC to represent it more comprehensively. A
four-layer CNN network named TSCNN-DS was designed for the
classification process. The proposed model achieved 97.2 % accu-
racy on the US8K dataset. In the study by Mushtaq et al. [43], it
was emphasized that the overlapping of sounds increases the com-
plexity with the inclusion of many sound sources in the environ-
ment in the distance between the acoustic source and the
microphone. Due to this complexity, a deep convolutional neural
network (DCNN) is used with the validated audio properties of
the ESC. In the study, three sound feature extraction techniques,
Mel spectrogram, Mel Frequency Cepstral Coefficient, and Log-
Mel were considered for the ESC-10, ESC-50, and US8K datasets.
The obtained accuracy values were 94.94 %, 89.28 %, and 95.37 %
for the ESC-10, ESC-50 and US8K, respectively. In the study con-
ducted by Mushtaq et al. [44], an approach on spectral images in
classification with CNN was introduced using the method of direct
data augmentation of audio data. In the presented approach, the
Mel spectrogram feature is used. Randomly selected 7- or 9-layer
CNN models and ESC-10, ESC-50 and US8K data were used. It
was emphasized that effective and high accuracy was achieved
by using the meaningful data augmentation method directly over
the voice. Among the models used, accuracy rates of 99.04 % were
achieved for ResNet-52, ESC-10, 99.49 % for US8K datasets, and
97.57 % for DenseNet-161, ESC-50.
In the study by Chen et al. [45], the ESC problem was handled
with extended CNN. It has been stated that using this structure
results in higher classification accuracy than the max-pooling pro-
cess used after the convolution layer. At the same time, the effect
Fig. 2. Flow chart of particle swarm optimization algorithm.
of different expansion ratios and the number of convolution layers

Fig. 1. The basic structure of CNN architectures in deep learning.

2
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

on the results was investigated. In the ESC problem, the expanded
CNN achieved better results than the maximum pooling CNN;
however, it has been stated that increasing the number and ratio
of filters negatively affects the classification accuracy.
In the study by Abdoli et al. [46], a one-dimensional (1D) CNN
network was used for ESC classification. Frames are received over
the audio signal for the input data. Because of the experimental
studies, an average accuracy value of 89 % was obtained on the
US8K data set. It was stated that the use of raw input data showed
the best performance among the end-to-end studies. In addition, it
was stated that the proposed method has fewer parameters than
other models in the literature.
Conditional Neural Network (CLNN) and its extension Masked
ConditionaL Neural Network (MCLNN) have been proposed by
Medhat et al. [47], In the study, the time-frequency representation
of the sound was modeled by making use of the natural state of the
sound. With the proposed MCLNN, the classification process of dif-
ferent music and ESCs was achieved. MCLNN classification accu-

Fig. 3. An example representation of records of classes in the ESC10 dataset.

Table 1
Categories in the ESC50 dataset and the names of the classes in each category.

Animals Natural sound spaces & water sounds Human, Non-speech sounds Interior/domestic sounds Exterior/urban noises
Dog Rain Crying baby Door knock Helicopter
Rooster Sea waves Sneezing Mouse click Chainsaw
Pig Crackling fire Clapping Keyboard typing Siren
Cow Crickets Breathing Door, wood creaks Car horn
Frog Chirping birds Coughing Can opening Engine
Cat Water drops Footsteps Washing machine Train
Hen Wind Laughing Vacuum cleaner Church bells
Insects (flying) Pouring water Brushing teeth Clock alarm Airplane
Sheep Toilet flush Snoring Clock tick Fireworks
Crow Thunderstorm Drinking, sipping Glass breaking Hand saw

3
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Table 2
Number of records used for training, verification, and testing in the US8K data set.

Class Number of images Train Validation Test

Air conditioner 1000 600 200 200
Car horn 429 257 86 86
Children playing 1000 600 200 200
Dog bark 1000 600 200 200
Drilling 1000 600 200 200
Engine idling 1000 600 200 200
Gun shot 374 224 75 75
Jackhammer 1000 600 200 200
Siren 929 557 186 186
Street music 1000 600 200 200

Fig. 4. An example representation of records of classes in the US8K dataset.

racy outperforms CNN-based models. In the study conducted by
Zhang et al. [48], the effects of the sizes and activation functions
of the filters in the convolutional layer in CNNs on the ESC were
investigated. For this purpose, a Dilated CNN-based model (D-
CNN-ESC) is proposed. The proposed system has been applied to
three different sound data sets. The percent less error value was
obtained in the US8K dataset compared to other methods.
In the study by Lim et al. [49], a CNN-based method was pro-
posed for the classification of sound events. The proposed method
classifies 30 different sound events on different datasets with an

4
 _
Ö. Inik
Fig. 5. The architecture of the proposed method.
Fig. 6. Sound (up) and transformed image according to a scalogram (down) of classes in the US8K data set.
accuracy rate of 81.5 %. In Akbal’s study [50], a stable feature
extraction method is presented to determine the location of the
activity from environmental sounds. This method consists of three
basic stages, which include feature creation, selection, and classifi-
cation. In this study, the proposed method is applied to the ESC-10
data set and the classification of the sounds in the data set is pro-
vided. With this proposed method, an accuracy rate of 90.25 % was
obtained. Although deep learning models achieve high accuracies,
the models can be misled by noise added to the data. This has been
addressed in a detailed study by Tripathi and Mishra [52]. In their
study, the effect of attacks on ESC data classified with CNN models
was investigated. They also created two datasets that will be
benchmarks in these studies. Finally, in Tuncer et al. [53] ESC clas-
sification was performed by using spiral patterns and two-
dimensional maximum, minimum, median, and mean (2D-M4)
5
Applied Acoustics 202 (2023) 109168
pooling methods. This impressive study has been applied to the
ESC-10 and ESC-50 datasets, and high accuracy has been achieved.
1.2. Motivation
In general, it has been observed that the success rates of ESC
obtained with deep learning models have better results than other
artificial intelligence methods. The main reason for this can be
summarized as automatic feature discovery in deep learning mod-
els. However, many parameters need to be adjusted in the design
of deep learning models. These are divided into two parts: opti-
mization parameters and model design parameters. This situation
can be seen as both an advantage and a disadvantage. As an advan-
tage, it allows different models to be designed by different
researchers for the solution of a problem. On the other hand, the
 _
Ö. Inik
Fig. 7. The Flow chart of PSO used in CNN optimization.
fact that it is almost impossible to design the best model for a
problem can be seen as a disadvantage. A lot of work has been done
in recent years to overcome this disadvantage of CNN models.
These studies used different algorithms for the optimization of
CNN parameters. Optimization methods such as the genetic algo-
rithm, the evolutionary algorithm, and particle swarm optimiza-
tion were used for the parameter optimization of CNN models.
Satisfactory results have been obtained for the optimization of
CNN parameters [54–56]. Recently, CNN models [2,36–38,41–46,
48,49,57] have become used for ESC tasks. However, no studies
have been found so far regarding the most appropriate layer struc-
ture and parameter value for CNN to be used in ESC tasks. The main
Fig. 8. Representation of CNN layers as particles.
6
Applied Acoustics 202 (2023) 109168
motivation of this work is to perform the optimization of CNN
hyper-parameters, which gives the highest accuracy for the classi-
fication of urban sounds.
1.3. Contributions
Signal processing methods, classical machine learning methods,
and deep learning models are used for the classification of urban
sounds. The contribution of this study in addition to these studies
is given below.
1. Optimization of hyperparameters of CNN used for the classifica-
tion of urban sounds was carried out for the first time in this
study.
2. The PSO [58] algorithm was adapted for CNN parameter opti-
mization. Thus, the CNN parameters were optimized to be used
directly in the PSO algorithm without converting them into any
numerical values.
3. Unlike any transfer learning or CNN models designed in differ-
ent architectures, a basic CNN model that works with the high-
est accuracy was obtained.
4. With the parameter optimization performed in the study, a CNN
model was automatically obtained for the classification of envi-
ronmental sound data in image format. Thus, it is possible to
obtain an automatic CNN for those who want to work on any
data set in image format but have no knowledge of CNN design
or are beginners.
1.4. Paper structure
This paper is organized as follows. In section 2, information
about the CNN, PSO, and ESC data sets is given. In section 3, infor-
mation about the proposed method is given. This section describes
how to optimize CNN models with PSO. In addition, CNN models
that give the highest classification accuracy for each data set are
given with the proposed method. Experimental studies have been
conducted in section 4. Finally, the conclusion is explained in sec-
tion 5.
2. Background
2.1. Convolutional neural networks
Convolutional neural networks (CNN) are accepted as the main
architecture of deep learning, because of discovering its unique
features by performing the learning process on raw data. It exhibits
high performance in problems such as classification, identification,
and segmentation. For this reason, it has become widespread in
many fields such as engineering, medicine, and the defense indus-
try. It has become even more popular with its use with big data,
especially thanks to automatic feature discovery [59]. CNNs are
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Fig. 9. Calculation of Gbest-Particle or Pbest-Particle.

Fig. 10. Calculation of velocity.

Fig. 11. Update of particle.

7
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Table 3 ers. In this way, countless trial areas were left for the researcher to
Particle swarm optimization initial values. reach the highest performance ratio in the model he will create.
Name of Parameters Value The basic structure of the CNN is given in Fig. 1.
Number of iterations 3 Looking at the CNN architecture, while the section where the
Swarm size 50 input layer, convolution layer, and pooling layers are used for fea-
c 0.8 ture extraction, classification is performed in the section created by
the fully connected layers. In the working principle of CNNs, the
appropriate data received in the input layer is passed through
blocks consisting of convolution and pooling layers and transferred
to the last fully connected layer. After this layer, output data is
Table 4
Parameters values of CNN models architecture.
obtained. This data is compared with the desired result, and the
difference is accepted as an error. To minimize this error, the
Name of Parameters Value weights are updated using a gradient-based backpropagation algo-
Minimum number of layers 3 rithm. Network training is terminated when the desired epoch
Maximum number of layers 15 value is reached [60].
Minimum number of filters in a convolutional layer 16
Maximum number of filters in a convolutional layer 256
Minimum size of a convolutional kernel 2
2.2. Particle swarm optimization
Maximum size of a convolutional kernel 11
Minimum size of kernel a pooling layer 2
Maximum size of kernel a pooling layer 7 Particle swarm optimization (PSO), developed by Kennedy
Minimum size of stride a pooling layer 2 Eberhart in 1995, was inspired by the movements of herd animals
Maximum size of stride a pooling layer 7 in search of food [58]. The population created in this algorithm is
Minimum number of neurons in a fully connected layer 10
Maximum number of neurons in a fully connected layer 1024
called the swarm, and each individual in the population is called
a particle. Each particle searches for the best solution, and this is
called the ‘Pbest’. The search for the best solution made by the
whole population is called the ‘Gbest’. The flow diagram of the
Table 5 algorithm is given in Fig. 2. After the initial parameters are defined
CNN training parameters. in Fig. 2, the particles in the population are randomly distributed to
Parameters Value create a swarm. The Fitness function value is calculated according
to the positions of the particles created first, and then the Pbest
Optimizer SGDM
Epoch used to obtain the best CNN 10 and Gbest values are updated. The velocities and positions of the
Epoch for training the best CNN 50 particles are then recalculated. The position calculation of each
Dropout rate 0.5 particle is made according to Equation (1). The velocity calculation
Mini batch size 256 of each particle in Equation (1) is made according to Equation (2).
Initial learning rate 0.001

xiðtþ1Þ ¼ xiðtÞ þ v iðtþ1Þ ð1Þ

v iðt þ 1Þ ¼ w  v iðtÞ þ C 1  Rand1  Pbesti  xiðtÞ þ C 2


deep networks that consist of multiple layers in succession and  Rand2  Gbesti  xiðtÞ ð2Þ
extract feature maps of objects during network training. In addi-
tion to the advantages of CNNs in extracting the feature maps where i is the number of particles, t is the number of iterations, v is
themselves, it also left the researcher free in the number of layers particle velocity, x is particle position, Rand1 and Rand2 are random
in the model to be used and the parameters to be used in these lay- functions, Pbest is the best position ever visited by particle i, Gbest is

Fig. 12. Convergence graphs of the CNN-ESC10 model during the training.

8
 _
Ö. Inik
Fig. 13. Convergence graphs of the CNN-ESC50 model during the training.
Fig. 14. Convergence graphs of the CNN-US8K model during the training.
the best position discovered so far, and w, C1, and C2 are positive
constants (the inertia weight and acceleration coefficients).
2.3. Data sets
Three different data sets have been used extensively for the
classification of environmental sounds in recent years. These data-
sets are ESC10, ESC50, and UrbanSaund8K, respectively. Informa-
tion about these datasets is given below.
2.3.1. ESC-10 data set
The ESC-10 dataset was created by Karol J. Piczak [24]. There are
ten classes in the dataset; these classes are ‘Dog bark’, ‘Rain’, ‘Sea
waves’, ‘Baby cry’, ‘Clock tick’, ‘Person sneeze’, ‘Helicopter’, ‘Chain-
saw’, ‘Rooster’, and ‘Fire crackling’. Each recording consists of an
average of 5 s long. In total, there is an average of 40 audio record-
ings for each class. An example representation of the sounds
belonging to each class in the dataset is given in Fig. 3. This dataset
is a subset of the ESC-50 dataset.
2.3.2. ESC-50 data set
The ESC-50 dataset is an enhanced version of the ESC-10 data-
set. There are 50 classes in this dataset. Each class consists of 40
9
Applied Acoustics 202 (2023) 109168
recordings of 5 s in length. It has 2000 sound clips from urban set-
tings. These classes consist of the main categories: ‘Animals, natu-
ral soundscapes & water sounds’, ‘Human non-speech sounds’,
‘Interior/domestic sounds’, and ‘Exterior/urban noises’. The names
of each class under these main categories are given in Table 1.
2.3.3. UrbanSound8K
UrbanSound8K (US8K) is a dataset consisting of 8732 labeled
sounds prepared by Salamon et al. [25]. Each record in the dataset
is approximately 4 s long. The dataset has 10 classes. These classes
are: ‘air conditioner’, ‘car horn’, ‘children playing’, ‘dog bark’, ‘dril-
ling’, ‘engine idling’, ‘gunshot’, ‘jackhammer’, ‘siren’, and ‘street
music’. The number of records for each class in this data set is given
in Table 2. An example representation of the sounds belonging to
each class in this dataset is given in Fig. 4.
3. Proposed method
The architecture of the proposed method is given in Fig. 5. In the
proposed method, firstly, the audio data is converted to image for-
mat using the scalogram method. Secondly, the CNN model that
gave the best results for the data set was obtained with the PSO
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Fig. 15. Confusion matrix giving the average accuracy value of the CNN-ESC10 model.

Fig. 16. Confusion matrix giving the average accuracy of the CNN-ESC 50 model.

algorithm. The optimization of hyper-parameters in the CNN
model is detailed in Section 3.2.
3.1. Data set preprocessing
In the study, sounds were converted from signal to image for-
mat using the scalogram method. The scalogram is the absolute
value of the continuous wavelet transform of a signal plotted as
a function of time and frequency. The Wavelet Toolbox of Matlab
R2020b software was used for the conversion process. As an exam-
ple, the records converted from sound to image in the US8K dataset
are given in Fig. 6.
Since the models obtained in experimental studies are deeper, it
has been understood that more data is needed for the training of
these models. For this reason, the models were trained with both
original data sets and augmented data sets. Translation and flip
methods were used for data augmentation. By using these methods
in vertical, horizontal, and vertical horizontal directions, the data
set was increased three times in each method and the data sets
were increased 6 times in total.

10
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Fig. 17. Average confusion matrix of model ESC-US8K.

Fig. 18. Boxplot of the CNN-ESC10, CNN-ESC50, and ESC-US8K models.

3.2. Optimization of CNN Hyper-parameters. any task. However, this is not an easy process because of the num-
ber of parameters to be adjusted, and the computational cost is
CNN parameter optimization aims to perform the parameter very high. For this reason, it is necessary to use optimization algo-
selection for the CNN model that gives the highest accuracy for rithms that will result in lower iterations. In this study, the PSO

11
 _
Ö. Inik
Table 6
Performance values obtained with the proposed models.
Proposed Model Data Augmentation Mean Accuracy
CNN-ESC10 Yes 98.64
CNN-ESC10 No 88.50
CNN-ESC50 Yes 96.77
CNN-ESC50 No 74.85
CNN– US8K Yes 98.45
CNN– US8K No 91.17
algorithm is used to obtain the CNN model that will give the high-
est accuracy for the classification of urban sounds. For optimization
with PSO, first, the calculation made in Equation-1 and Equation-2
should be adapted to the CNN structure. For this reason, a new
method has been developed for the representation of CNN layer
structures, similar to the studies in [61,62]. In the proposed
method, CNN layers are used directly without converting any
numerical values. The flow chart of the updated PSO algorithm is
given in Fig. 7.
In Fig. 7, the first step in the flow chart is the creation of the par-
ticles. As shown in Fig. 8, each particle was formed to represent a
CNN model. The best model is obtained after the created CNN mod-
els are trained with the data set. At this stage, Gbest represents the
best-performing CNN model and Pbest represents the first CNN
model. In the next step of the flow diagram, the velocity is calcu-
lated according to Equation (3).
 The parameters used for the training of each CNN model are
Pbest  Pif ðr > r v Þ
Velocity ¼ ð3Þ
Gbest  Pelse
The r in the equation refers to the randomly generated number
in the range 0–1. The rv represents the reference value and deter-
mines which layer will be taken in the comparison. In the algo-
rithm, the Gbest – P or Pbest – P difference is calculated as shown
in Fig. 9. In Fig. 9, by comparing each convolution and pooling
layer, the same layers take the letter ‘‘O” and in fully connected
layers O (fc). If the values in the layer are different, the Pbest and
Gbest values are taken into account. An example of velocity calcu-
lation according to Equation (3) is given in Fig. 10. As seen in
Fig. 10, the rv value was determined as 0.8. For the first layer, r
was generated as 0.74. Hence, the first layer of velocity (Gbest –
Particle) is taken. In this way, the velocity value is obtained after
comparing all layers according to the randomly generated r-
value. After the velocity is obtained, the particles are updated as
shown in Fig. 11. In Fig. 11, if the velocity layer is 0, the particle
layer is considered. If the velocity and the particle layers are differ-
ent, the velocity layer is taken.
4. Experimental studies
In this section, training of CNN models obtained with PSO and
test results obtained by these models are given. The success values
obtained were compared with the studies in the literature. In addi-
tion, both the values of the parameters used in the design of the
CNN models and the initial values of the parameters used in the
PSO are given. Experimental studies have been done on a computer
with IntelÒ CoreTM i9-7900X 3.30 GHz  20 processor, 64 GB RAM
and 2  GeForce RTX2080Ti graphic card. Matlab R2021b 64bit
(win64) was used as the software platform.
4.1. Parameters used in the proposed method
The parameters used for Adapted-PSO are given in Table 3. As
seen in Table 3, the number of swarms was determined as 50. Each
12
Applied Acoustics 202 (2023) 109168
Min Accuracy Max Accuracy Standard Deviation
99.64 97.86 0.59
95.00 80.00 4.64
97.36 96.36 0.34
81.00 71.00 2.82
98.64 98.18 0.13
92.90 88.89 1.03
swarm represents a CNN layer. Therefore, 50 different CNN layers
are created in each iteration. Since the convergence speed is very
high in PSO, the number of iterations was chosen as low as 3.
The c value in the table represents the rv value in Eq. (3). An exam-
ple usage of this value in the proposed method is given in Fig. 10.
The parameter values used for the optimization of the CNN
model are given in Table 4. In particular, the number of parameters
has been kept wide to find the most suitable model by searching
the CNN models in a wide solution space. The maximum number
of layers of the models has been selected as 15. The minimum
number of layers is set to three, and one of them must be a fully
connected layer. In addition, the fully connected layer must always
be placed last in the sequence of layers. Layers represent only con-
volution, pooling, and fully connected layer numbers. The ReLu lay-
ers after the convolution layers and the DropOut layers after the FC
layers are added automatically.
given in Table 5. In the study, there are two types of epoch values:
the epoch value of the CNNs trained during optimization and the
epoch value of the best CNN found after optimization is completed.
In the first, the epoch value was chosen as 10 to find the CNN
model with the best performance in the training phase. Indeed,
10 epochs are a low value for training a CNN model. However, it
is acceptable for all candidate CNN models to have an epoch value
of 10 and to choose the best among them. In the second, the best
candidate is trained with the final training parameters and epoch
value of 50.
4.2. Best CNN models obtained with adapted PSO
The best CNN models obtained for the ESC10, ESC50, and US8K
datasets are named CNN-ESC10, CNN-ESC50, and CNN-US8K,
respectively. The layer architecture and parameters of these mod-
els are given Appendix A. The CNN-ESC10 model consists of a total
of 22 layers, including 4 convolutional, 3 pooling, and 2 fully con-
nected layers. The CNN-ESC50 model consists of a total of 24 lay-
ers, including 5 convolutional, 2 pooling, and 2 fully connected
layers. The CNN-US8K model consists of a total of 18 layers, includ-
ing 4 convolutional, 3 pooling, and 2 fully connected layers. The
CNN-ESC50 model consists of a total of 24 layers, including 5 con-
volutional, 2 pooling, and 2 fully connected layers.
4.3. Training and results of CNN models
The training process for the CNN models was performed with
10-fold cross-validation. The convergence graphs of the CNN-
ESC10, CNN-ESC50, and CNN-US8K during the training phase are
given in Fig. 12, Fig. 13, and Fig. 14, respectively. When all three
figures are examined, it can be seen that the accuracy and valida-
tion graphs and the error and validation error graphs are very close
to each other; therefore, the models do not overfit when training.
In addition, CNN-ESC10 converges faster than other models.
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

The confusion matrix giving the average accuracy of the CNN-
ESC10 model is given in Fig. 15. Looking at Fig. 15, the model
labeled the Rooster class with the highest accuracy and the Dog
bark class with the lowest accuracy. In addition, the confusion
matrices obtained by this model for each fold are given in Appen-
dix B. The highest accuracy rate was obtained in Fold 9, and the
lowest accuracy rate was obtained in Fold 4.
The confusion matrix giving the average accuracy of the CNN-
ESC50 model is given in Fig. 16. Looking at Fig. 16, the model labels
26 classes with 100 % accuracy. Confusion matrices obtained by the
model for each fold are given in Appendix C.
The confusion matrix giving the average accuracy of the CNN-
US8K model during the test phase is given in Fig. 17. Looking at
Fig. 17, the class labeled with the highest accuracy by the model
is Air conditioner and the classes labeled with the lowest accuracy
are Car horns and Street music. In addition, the confusion matrices
obtained by this model for each fold are given in Appendix D.
Boxplots of all three models are given in Fig. 18. As seen in
Fig. 18, CNN-ESC10 showed the best performance, while the
CNN-ESC10 model showed the lowest performance. In addition,
the CNN-US8K model showed the lowest standard deviation.
The performance values of the models obtained in this study on
the data sets are presented in Table 6. In the table, the results
obtained with or without data augmentation are given for each
data set. The average accuracy values obtained in the ESC-10,
ESC-50, and US8K datasets are 88.50, 74.85, and 91.17 % without
data augmentation, and 98.64, 96.77, and 98.45 % with data aug-
mentation, respectively. It was seen that the models obtained
higher values with data augmentation. The highest accuracy value
between normal data and augmented data was seen in the ESC-50
data set. The main reason for this is that there are 50 classes in this
dataset and less data for training because the model could not find
enough datasets to update its weights. With the increase of the
data, the success values for 50 classes increased more.
4.4. Comparison with other studies
Different studies based on deep learning have been conducted
on the ESC data set. The comparison of the average accuracy value
obtained with the proposed method with other studies is given in
Table 7. The ‘‘–” character in Table 7 means that the relevant
method was not used on that data set. It has been observed that
the proposed CNN models have higher success than the baseline
models [24,25] in all data sets. Compared to all studies in the table,
it outperformed all studies except for one [44]. Since transfer
learning is used in this study, it is expected to obtain a higher accu-
racy rate because, in transfer learning, the models were applied to
urban sounds after being trained on a different data set consisting
of millions of data. The proposed models yielded better results
than human performance in the ESC-10 and ESC-50 datasets.
4.5. Discussion
Classification of urban sounds with artificial intelligence meth-
ods has increased in recent years. Due to the high success of deep
learning models in other data types, it has been used frequently in
urban sounds. It is still being investigated by many researchers to
obtain CNN models that perform better in the classification of
urban sounds. However, in the studies carried out so far, the mod-
els were designed manually. Therefore, it is almost impossible to
obtain CNN models that work with the highest accuracy because
the number of parameters that need to be set in a CNN is very
large. For example, just the order of the layers can be in a million
different ways. In addition, considering that the parameter values

Table 7
Comparison of the accuracy value obtained by the proposed method with other methods’ accuracy (%).

Year-Authors [References] Method US8K ESC10 ESC50

2014- Salamon et al. [25] SVM (Baseline) 68.00 – –
2015-Piczak [24] Random Forest Ensemble (Baseline) – 72.70 44.30
2015- Piczak [24] Human Performance – 95.70 81.30
2017- Boddapati et al. [40] AlexNet, GoogLeNet, CRNN 93.00 91.00 73.00
2020- Inik and Seker [63] CNN 82.45 – –
2017- X Zhang et al. [48] D-CNN (Activation func. = LeakyReLU) 81.90 – 68.10
2017- X Zhang et al. [48] D-CNN (Activation func. = PReLU) 81.40 – 66.20
2017- X Zhang et al. [48] D-CNN (Activation func. = ReLU) 81.20 – 67.14
2016- Yusuf Aytar et al. [64] SoundNet – 92.20 74.20
2016- Salamon and Bello [37] DCNN + data augmentation 79.00 – –
2017- X Zhang et al. [48] D-CNN (Activation func. = ELU) 78.90 – 68.00
2020-Demir et al. [2] Pyramid-Combined CNN 78.14 94.80 81.40
2019-Chen et al. [45] Dilated CNN 78.00 – –
2017-Ye et al. [65] feature learning 77.36 – –
2017- X Zhang et al. [48] D-CNN(Activation func. = Softplus) 73.70 – 53.00
2017- Dai et al. [66] CNN 71,68 – –
2018- Pons and Serra[67] VGG and SVMs 70.74 – –
2020-Akbal [50] SVM – 90.25 –
2021-Mushtaq et. al. [44] ResNet-152, DenseNet-161 99.49 99.04 97.57
2018- Zhu B et al. [68] WaveMsNet – 93.75 79.10
2019- Xinyu Li et al. [69] Multi-Stream CNN – 93.70 83.50
2021-S. Luz et al.[70] Handcrafted + Deep 96.80 – 86.20
2021- Tripathi and Mishra [71] self-supervised learning (SSL) – 91.67 –
2021- Zhang et al. [72] RNN – 93.70 86.10
2021- Tripathi and Mishra [73] Residual Network – 92.00 –
2022- Zhang et al. [74] PSO + ensemble CBiLSTM – 93.00 –
2020-Medhat et al.[47] Masked Conditional Neural Networks 74.22 85.25 66.60
Proposed CNN optimization with PSO 91.17 88.50 74.85
Proposed (Data Augmentation) CNN optimization with PSO 98.45 98.64 96.77

13
 _
Ö. Inik
in each layer need to be adjusted, it is understood how wide the
solution space of the parameters needs to be adjusted for a CNN.
For this reason, obtaining the ideal solution in this wide solution
space can only be achieved with optimization algorithms. There-
fore, in this study, the CNN model with the best performance
was sought using the PSO algorithm for the classification of urban
sounds. The most important reason for using the PSO algorithm is
the high convergence speed of this algorithm. Therefore, it is the
ideal solution for CNNs with high computational costs, as it gives
better results in lower iterations.
One of the most important issues for CNN parameter optimiza-
tion is the representativeness of these models for parameter opti-
mization. In this study, a unique representation method was
created by modifying the PSO algorithm. Thus, unlike some studies
[54,75,76], the CNN parameters were optimized to be used directly
in the PSO algorithm without converting them to any numerical
value. Compared to other studies [37,45,48,66] using pure CNN
in the classification of urban sounds, CNN models obtained with
PSO performed much better. It has been seen that the CNN models
obtained with PSO are deeper. It was understood that the weights
of these deep networks would need more data to reach the opti-
mum value, and the models were trained by generating synthetic
data in the data sets. Especially with the high number of classes
in the ESC-50 data set and the low training data in each class,
the weights of the CNN-ESC50 model could not update sufficiently
with the raw data. When this data set was increased, and the
model was trained again, the accuracy rate increased from
74.85 % to 96.77 %. This result suggests that CNN models obtained
for ESC will also achieve higher accuracy rates in ESC classification
after training on a larger data set. Therefore, when the proposed
models were trained only with ESC sounds, they gave lower results
than their competitors [44] in all three datasets. These cutting-
edge models have achieved high success rates with transfer
learning.
In addition, some ensemble-based methods [70,74], which are
made by combining different features, achieved better results
when compared with the proposed models without data augmen-
tation because these methods used machine learning and hand-
crafted features in addition to the features found by CNNs, which
can be predicted to achieve better results.
In future studies, it is thought that more effective transfer learn-
ing studies will be carried out to use PSO-based CNN models
obtained on lower datasets with less computational cost in training
processes in large datasets.
Fig. A1. Layer architecture of the CNN-ESC10 model.
14
Applied Acoustics 202 (2023) 109168
5. Conclusion
In the study, the PSO algorithm was used to obtain the highest
CNN model on the ESC-10, ESC-50, and Urbansound8k data sets,
which are accepted as state of art in the classification of urban
sounds. In the proposed method, the PSO algorithm was modified
to optimize the CNN parameter. The best models for the ESC-10,
ESC-50, and Urbansound8k datasets were CNN-ESC10, CNN-
ESC50, and CNN-US8K, with 88.50, 74.85, and 91.17 % accuracy
rates, respectively. These results were obtained from raw data.
The data sets were increased seven times with data augmentation
techniques, and the same models were trained again. The results
were obtained with 98.64, 96.77, and 98.45 % accuracy, respec-
tively. When these results are compared with competing studies,
the CNN models obtained for the classification of urban sounds
were very successful.
CRediT authorship contribution statement
_
Özkan Inik: Conceptualization, Data curation, Formal analysis,
Funding acquisition, Investigation, Methodology, Project adminis-
tration, Resources, Software, Validation, Visualization, Writing –
original draft, Writing – review & editing.
Data availability
Data will be made available on request.
Declaration of Competing Interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
Appendix A. Layer architecture and parameters of CNN-ESC10,
CNN-ESC10, and US8K models
Fig. A1.
Table A1.
Fig. A2.
Table A2.
Fig. A3.
Table A3.
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Table A1
Parameter values in each layer of the CNN-ESC10 model.

No Layer Activations Number of Filter Filter Size Stride Total Parameter

1 Input Layer 2242243 – – – –
2 Conv. 224224140 140 66 11 663140
3 Batch Norm. 224224140 – – – 111140
4 ReLu Layer 224224140 – – – –
5 Avr. Pooling 4545140 – 44 55 –
6 Conv. Layer 4545205 205 1010 11 1010140205
7 Batch Norm. 4545205 – – – 11205
8 ReLu 4545205 – – – –
9 Max. Pooling 2121205 – 44 22 –
10 Conv. 2121183 183 99 11 99205183
11 Batch Norm 2121183 11183
12 ReLu 2121183 – – – –
13 Max. Pooling 33183 – 55 77 –
14 Conv. 33103 103 1010 11 1010183103
15 Batch Norm 33103 – – – 11103
16 ReLu 33103 – – – –
17 Fully Connected 11763 – – – 763927
18 ReLu 11763
19 DropOut 11763 – – – –
20 Fully Connected 1110 – – – 10763
21 Softmax 1110 – – – –
22 Classification 1110 – – – –

Fig. A2. Layer architecture of the CNN-ESC50 model.

Table A2
Parameter values in each layer of the CNN-ESC50 model.

No Layer Activations Number of Filter Filter Size Stride Total Parameter

1 Input 2242243 – – – –
2 Conv 224224109 109 22 11 223109
3 Batch. Norm. 224224109 – – – 11109
4 ReLu 224224109 – – – –
5 Avg. Pooling 5656109 – 44 44 –
6 Conv. 5656203 203 22 11 22109203
7 Batch. Norm. 5656203 – – – 11203
8 ReLu 5656203 – – – –
9 Max. Pooling 1818203 – 44 33 –
10 Conv. 1818181 181 33 11 33203181
11 Batch. Norm. 1818181 11181
12 ReLu 1818181 – – – –
13 Conv. 1818210 210 44 11 44181210
14 Batch. Norm. 1818210 – – – 11210
15 ReLu 1818210 – – – –
16 Conv. 1818169 169 44 11 44210169
17 Batch. Norm. 1818169 – – – 11169
18 Relu 1818169 – – – –
19 Fully Connected 11850 – – – 85054756
20 ReLu 11850 – – –
21 DropOut 11850 – – – –
22 Fully Connected 1150 – – – 50850
23 Softmax 1150 – – – –
24 Classification 1150 – – – –

15
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Fig. A3. Layer architecture of the CNN-US8K model.

Table A3
Parameter values in each layer of the CNN-US8K model.

No Layer Activations Number of Filter Filter Size Stride Total Parameter

1 Input 2242243 – – – –
2 Conv. 22422418 18 88 11 88318
3 ReLu 22422418 – – – –
4 Avrg. Pooling 373718 – 55 66 –
5 Conv. 3737214 214 88 11 8818214
6 ReLu 3737214 – – – –
7 Avrg. Pooling 1111214 – 55 33 –
8 Conv. 1111249 249 44 11 44214249
9 ReLu 1111249 – – – –
10 Conv. 1111229 229 33 11 33249229
11 ReLu 1111229 – – – –
12 Avrg. Pooling 33229 – 33 44 –
13 Fully Connected 11714 – – – 7142061
14 DroupOut 11714 – – – –
15 ReLu 11714 – – – –
16 Fully Connected 1110 – – – 10714
17 Softmax 1110 – – – –
18 Classification 1110 – – – –

16
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Appendix B. Confusion matrices obtained by the CNN-ESC10 model. The highest accuracy was achieved in Fold-9

17
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Appendix C. Confusion matrices obtained by the CNN-ESC50 model. The highest accuracy was achieved in Fold-3

18
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

19
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

20
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

21
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

22
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

Appendix D. Confusion matrices obtained by the ESC-US8K model. The highest accuracy was achieved in Fold-6

23
 _
Ö. Inik Applied Acoustics 202 (2023) 109168

References [31] J. T. Geiger and K. Helwani, ‘‘Improving event detection for audio surveillance
using gabor filterbank features,” in 2015 23rd European Signal Processing
Conference (EUSIPCO), 2015, pp. 714-718.
[1] Chu S, Narayanan S, Kuo C-C-J. Environmental sound recognition with time–
[32] Mulimani M, Koolagudi SG. Segmentation and characterization of acoustic
frequency audio features. IEEE Trans Audio Speech Lang Process
event spectrograms using singular value decomposition. Expert Syst Appl
2009;17:1142–58.
2019;120:413–25.
[2] Demir F, Turkoglu M, Aslan M, Sengur A. A new pyramidal concatenated CNN
[33] Xie J, Zhu M. Investigation of acoustic and visual features for acoustic scene
approach for environmental sound classification. Appl Acoust
classification. Expert Syst Appl 2019;126:20–9.
2020;170:107520.
[34] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘Imagenet classification with
[3] Aumond P, Lavandier C, Ribeiro C, Boix EG, Kambona K, D’Hondt E, et al. A
deep convolutional neural networks,” in Advances in neural information
study of the accuracy of mobile technology for measuring urban noise
processing systems, 2012, pp. 1097-1105.
pollution in large scale participatory sensing campaigns. Appl Acoust
[35] Deng J, Berg A, Satheesh S, Su H, Khosla A, Fei-Fei L. Imagenet large scale visual
2017;117:219–26.
recognition competition 2012 (ILSVRC2012). See net org/challenges/LSVRC
[4] Cao J, Cao M, Wang J, Yin C, Wang D, Vidal P-P. Urban noise recognition with
2012:41.
convolutional neural network. Multimed Tools Appl 2019;78:29021–41.
[36] K. J. Piczak, ‘‘Environmental sound classification with convolutional neural
[5] Radhakrishnan R, Divakaran A, Smaragdis A. Audio analysis for surveillance
networks,” in 2015 IEEE 25th International Workshop on Machine Learning for
applications. IEEE Workshop on Applications of Signal Processing to Audio and
Signal Processing (MLSP), 2015, pp. 1-6.
Acoustics 2005;2005:158–61.
[37] Salamon J, Bello JP. Deep convolutional neural networks and data
[6] Crocco M, Cristani M, Trucco A, Murino V. Audio surveillance: A systematic
augmentation for environmental sound classification. IEEE Signal Process
review. ACM Computing Surveys (CSUR) 2016;48:1–46.
Lett 2017;24:279–83.
[7] Laffitte P, Wang Y, Sodoyer D, Girin L. Assessing the performances of different
[38] N. Takahashi, M. Gygli, B. Pfister, and L. Van Gool, ‘‘Deep convolutional neural
neural network architectures for the detection of screams and shouts in public
networks and data augmentation for acoustic event detection,” arXiv preprint
transportation. Expert Syst Appl 2019;117:29–41.
arXiv:1604.07160, 2016.
[8] Heittola T, Mesaros A, Eronen A, Virtanen T. Audio context recognition using
[39] Y. Tokozume, Y. Ushiku, and T. Harada, ‘‘Learning from between-class
audio event histograms. European Signal Processing Conference 2010:1272–6.
examples for deep sound recognition,” arXiv preprint arXiv:1711.10282, 2017.
in 2010 18th.
[40] Boddapati V, Petef A, Rasmusson J, Lundberg L. Classifying environmental
[9] Xu M, Xu C, Duan L, Jin JS, Luo S. Audio keywords generation for sports video
sounds using image recognition networks. Procedia Comput Sci
analysis. ACM Transactions on Multimedia Computing, Communications, and
2017;112:2048–56.
Applications (TOMM) 2008;4:1–23.
[41] Li S, Yao Y, Hu J, Liu G, Yao X, Hu J. An ensemble stacked convolutional neural
[10] A. Waibel, H. Steusloff, and R. Stiefelhagen, ‘‘CHIL-Computers in the human
network model for environmental event sound recognition. Appl Sci
interaction loop. 5th Intern,” in Workshop on Image Analysis for Multimedia
2018;8:1152.
Interactive Services, 2004.
[42] Su Y, Zhang K, Wang J, Madani K. Environment sound classification using a
[11] D. P. Ellis and K. Lee, ‘‘Minimal-impact audio-based personal archives,” in
two-stream CNN based on decision-level fusion. Sensors 2019;19:1733.
Proceedings of the the 1st ACM workshop on Continuous archival and retrieval of
[43] Mushtaq Z, Su S-F. Environmental sound classification using a regularized
personal experiences, 2004, pp. 39-47.
deep convolutional neural network with data augmentation. Appl Acoust
[12] Eronen AJ, Peltonen VT, Tuomi JT, Klapuri AP, Fagerlund S, Sorsa T, et al. Audio-
2020;167:107389.
based context recognition. IEEE Trans Audio Speech Lang Process
[44] Mushtaq Z, Su S-F, Tran Q-V. Spectral images based environmental sound
2005;14:321–9.
classification using CNN with meaningful data augmentation. Appl Acoust
[13] Barchiesi D, Giannoulis D, Stowell D, Plumbley MD. Acoustic scene
2021;172:107581.
classification: Classifying environments from the sounds they produce. IEEE
[45] Chen Y, Guo Q, Liang X, Wang J, Qian Y. Environmental sound classification
Signal Process Mag 2015;32:16–34.
with dilated convolutions. Appl Acoust 2019;148:123–32.
[14] H. Li, S. Ishikawa, Q. Zhao, M. Ebana, H. Yamamoto, and J. Huang, ‘‘Robot
[46] Abdoli S, Cardinal P, Koerich AL. End-to-end environmental sound
navigation and sound based position identification,” in 2007 IEEE International
classification using a 1D convolutional neural network. Expert Syst Appl
Conference on Systems, Man and Cybernetics, 2007, pp. 2449-2454.
2019;136:252–63.
[15] Lyon RF. Machine hearing: An emerging field [exploratory dsp]. IEEE Signal
[47] Medhat F, Chesmore D, Robinson J. Masked Conditional Neural Networks for
Process Mag 2010;27:131–9.
sound classification. Appl Soft Comput 2020;90:106073.
[16] Chu S, Narayanan S, Kuo C-C-J, Mataric MJ. ‘‘Where am I? Scene recognition for
[48] X. Zhang, Y. Zou, and W. Shi, ‘‘Dilated convolution neural network with
mobile robots using audio features,” in. IEEE International conference on
LeakyReLU for environmental sound classification,” in 2017 22nd International
multimedia and expo 2006;2006:885–8.
Conference on Digital Signal Processing (DSP), 2017, pp. 1-5.
[17] J. Huang, ‘‘Spatial auditory processing for a hearing robot,” in Proceedings. IEEE
[49] M. Lim, D. Lee, H. Park, Y. Kang, J. Oh, J.-S. Park, et al., ‘‘Convolutional Neural
International Conference on Multimedia and Expo, 2002, pp. 253-256.
Network based Audio Event Classification,” KSII Transactions on Internet &
[18] Green M, Murphy D. Environmental sound monitoring using machine learning
Information Systems, vol. 12, 2018.
on mobile devices. Appl Acoust 2020;159:107041.
[50] Akbal E. An automated environmental sound classification methods based on
[19] P. Intani and T. Orachon, ‘‘Crime warning system using image and sound
statistical and textural feature. Appl Acoust 2020;167:107413.
processing,” in 2013 13th International Conference on Control, Automation and
[51] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for large-
Systems (ICCAS 2013), 2013, pp. 1751-1753.
scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[20] Torija AJ, Ruiz DP, Ramos-Ridao ÁF. A tool for urban soundscape evaluation
[52] Tripathi AM, Mishra A. Adv-ESC: Adversarial attack datasets for an
applying support vector machines for developing a soundscape classification
environmental sound classification. Appl Acoust 2022;185:108437.
model. Sci Total Environ 2014;482:440–51.
[53] Tuncer T, Subasi A, Ertam F, Dogan S. A novel spiral pattern and 2D M4 pooling
[21] Romero VP, Maffei L, Brambilla G, Ciaburro G. Modelling the soundscape
based environmental sound classification method. Appl Acoust
quality of urban waterfronts by artificial neural networks. Appl Acoust
2020;170:107508.
2016;111:121–8.
[54] Ma B, Li X, Xia Y, Zhang Y. Autonomous deep learning: A genetic DCNN
[22] Agha A, Ranjan R, Gan W-S. Noisy vehicle surveillance camera: A system to
designer for image classification. Neurocomputing 2020;379:152–61.
deter noisy vehicle in smart city. Appl Acoust 2017;117:236–45.
[55] Gonçalves CB, Souza JR, Fernandes H. CNN architecture optimization using bio-
[23] Ntalampiras S. Universal background modeling for acoustic surveillance of
inspired algorithms for breast cancer detection in infrared images. Comput
urban traffic. Digital Signal Process 2014;31:69–78.
Biol Med 2022;142:105205.
[24] K. J. Piczak, ‘‘ESC: Dataset for environmental sound classification,” in
[56] Singh P, Chaudhury S, Panigrahi BK. Hybrid MPSO-CNN: Multi-level particle
Proceedings of the 23rd ACM international conference on Multimedia, 2015, pp.
swarm optimized hyperparameters of convolutional neural network. Swarm
1015-1018.
Evol Comput 2021;63:100863.
[25] J. Salamon, C. Jacoby, and J. P. Bello, ‘‘A dataset and taxonomy for urban sound
[57] Z. Zhang, S. Xu, T. Qiao, S. Zhang, and S. Cao, ‘‘Attention based convolutional
research,” in Proceedings of the 22nd ACM international conference on
recurrent neural network for environmental sound classification,” in Chinese
Multimedia, 2014, pp. 1041-1044.
Conference on Pattern Recognition and Computer Vision (PRCV), 2019, pp. 261-
[26] Bisot V, Serizel R, Essid S, Richard G. Feature learning with matrix factorization
271.
applied to acoustic scene classification. IEEE/ACM Trans Audio Speech Lang
[58] J. Kennedy and R. Eberhart, ‘‘Particle swarm optimization (PSO),” in Proc. IEEE
Process 2017;25:1216–29.
International Conference on Neural Networks, Perth, Australia, 1995, pp. 1942-
[27] Stowell D, Giannoulis D, Benetos E, Lagrange M, Plumbley MD. Detection and
1948.
classification of acoustic scenes and events. IEEE Trans Multimedia
[59] Dev D. Deep Learning with Hadoop. Packt Publishing Ltd; 2017.
2015;17:1733–46. _ Ülker E. Derin Öğrenme ve Görüntü Analizinde Kullanılan Derin Ög
[60] Özkan I,
[28] Dhanalakshmi P, Palanivel S, Ramalingam V. Classification of audio signals
˘renme Modelleri. Gaziosmanpasßa Bilimsel Arasßtırma Dergisi 2017;6:85–104.
using AANN and GMM. Appl Soft Comput 2011;11:716–23.
[61] Junior FEF, Yen GG. Particle swarm optimization of deep neural networks
[29] Ludena-Choez J, Gallardo-Antolin A. Acoustic Event Classification using
architectures for image classification. Swarm Evol Comput 2019;49:62–74.
spectral band selection and Non-Negative Matrix Factorization-based
[62] Passricha V, Aggarwal RK. PSO-based optimized CNN for Hindi ASR. Int J
features. Expert Syst Appl 2016;46:77–86.
Speech Technol 2019;22:1123–33.
[30] J. Salamon and J. P. Bello, ‘‘Unsupervised feature learning for urban sound
[63] O. Inik and H. Seker, ‘‘CnnSound: Convolutional Neural Networks for the
classification,” in 2015 IEEE International Conference on Acoustics, Speech and
Classification of Environmental Sounds,” in 2020 The 4th International
Signal Processing (ICASSP), 2015, pp. 171-175.
Conference on Advances in Artificial Intelligence, 2020, pp. 79-84.

24
 _
Ö. Inik
[64] Y. Aytar, C. Vondrick, and A. Torralba, ‘‘Soundnet: Learning sound
representations from unlabeled video,” in Advances in neural information
processing systems, 2016, pp. 892-900.
[65] Ye J, Kobayashi T, Murakawa M. Urban sound event classification based on
local and global features aggregation. Appl Acoust 2017;117:246–56.
[66] W. Dai, C. Dai, S. Qu, J. Li, and S. Das, ‘‘Very deep convolutional neural networks
for raw waveforms,” in 2017 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), 2017, pp. 421-425.
[67] J. Pons and X. Serra, ‘‘Randomly weighted CNNs for (music) audio
classification,” in ICASSP 2019-2019 IEEE international conference on acoustics,
speech and signal processing (ICASSP), 2019, pp. 336-340.
[68] Zhu B, Wang C, Liu F, Lei J, Huang Z, Peng Y, et al. ‘‘Learning environmental
sounds with multi-scale convolutional neural network,” in. International Joint
Conference on Neural Networks (IJCNN) 2018;2018:1–8.
[69] X. Li, V. Chebiyyam, and K. Kirchhoff, ‘‘Multi-stream network with temporal
attention for environmental sound classification,” arXiv preprint
arXiv:1901.08608, 2019.
25
Applied Acoustics 202 (2023) 109168
[70] Luz JS, Oliveira MC, Araujo FH, Magalhães DM. Ensemble of handcrafted and
deep features for urban sound classification. Appl Acoust 2021;175:107819.
[71] Tripathi AM, Mishra A. Self-supervised learning for Environmental Sound
Classification. Appl Acoust 2021;182:108183.
[72] Zhang Z, Xu S, Zhang S, Qiao T, Cao S. Attention based convolutional recurrent
neural network for environmental sound classification. Neurocomputing
2021;453:896–903.
[73] Tripathi AM, Mishra A. Environment sound classification using an attention-
based residual neural network. Neurocomputing 2021;460:409–23.
[74] Zhang L, Lim CP, Yu Y, Jiang M. Sound classification using evolving ensemble
models and Particle Swarm Optimization. Appl Soft Comput
2022;116:108322.
[75] L. Xie and A. Yuille, ‘‘Genetic cnn,” in Proceedings of the IEEE international
conference on computer vision, 2017, pp. 1379-1388.
[76] Sinha T, Haidar A, Verma B. ‘‘Particle swarm optimization based approach for
finding optimal values of convolutional neural network parameters,” in. IEEE
congress on evolutionary computation (CEC) 2018;2018:1–6.