Audio Classification: Identification of Single and Multiple

Sounds in an Audio Clip

A Project Report

Submitted

In partial fulfillment of the requirements for the award of the degree

BACHELOR OF TECHNOLOGY
In
COMPUTER SCIENCE and ENGINEERING

By
K. Supraja 211FA04381
K. Thanmai Ganga Bhavani 211FA04385

Under the Guidance of

Dr. G. Balu Narasimha Rao
Assistant Professor, CSE

DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

VIGNAN'S FOUNDATION FOR SCIENCE, TECHNOLOGY & RESEARCH

(Deemed to be University)
Vadlamudi, Guntur -522213, INDIA.

May, 2025
 DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

CERTIFICATE

This is to certify that the project report entitled “Audio Classification: Identification of
Single and Multiple Sounds in an Audio Clip” has been submitted by K.
Supraja(211FA04381), K. Thanmai Ganga Bhavani(211FA04385) in partial fulfillment of
the requirements for the Major Project course, as part of the academic curriculum of the
B.Tech. CSE Program, Department of Computer Science and Engineering (CSE) at
VFSTR Deemed to be University.

Project Guide HoD, CSE

External Examiner

i
 DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

DECLARATION
We hereby declare that the project work entitled "Audio Classification: Identification of Single
and Multiple Sounds in an Audio Clip" submitted in partial fulfillment of the requirements for
the award of the degree of Bachelor of Technology (B.Tech) in Computer Science and
Engineering at VFSTR Deemed to be University is a record of my/our original work.

This project has been carried out under the supervision of the Department of Computer Science
and Engineering, VFSTR Deemed to be University. The work embodied in this thesis has not
been submitted previously, in part or full, to any other University or Institution for the award of
any degree or diploma.

I/We have duly acknowledged all sources of information and data used in the preparation of this
project report and shall abide by the principles of academic integrity and ethical guidelines.

By

K. Supraja (211FA04381)
K. Thanmai Ganga Bhavani (211FA04385)

Date:

ii
 ACKNOWLEDGEMENT

We take this opportunity to express our deep sense of gratitude to our Project Guide, Dr. G.
Balu Narasimha Rao for granting us permission to undertake this project and for his
unwavering support, valuable guidance, and constant encouragement throughout the duration
of our work.

We are sincerely thankful to Dr. D. Yakobu, Mr. P. VenkataRajulu, Mr. D.

BalaKotaiah, and Mr. G.Murali Project Coordinators, for their consistent assistance,
timely advice, and cooperation, which were instrumental in the successful completion
of our project.

It is our privilege to extend our heartfelt thanks to Dr. S. V. Phani Kumar, Head of the
Department, Computer Science and Engineering, VFSTR Deemed to be University, for
providing us the opportunity and necessary resources to carry out this project work. We would
also like to express our profound gratitude to Dr. K. V. Krishna Kishore, Dean, School of
Computing and Informatics, VFSTR Deemed to be University, for his encouragement and
for facilitating an environment conducive to research and innovation.

We extend our sincere appreciation to all the faculty members, programmers, and technical
staff of the Department of Computer Science and Engineering for their valuable support,
knowledge sharing, and assistance throughout our academic journey.

Finally, we are deeply thankful to our family members for their unconditional love, constant
support, and encouragement, which were crucial in sustaining our efforts and successfully
completing this project.

With Sincere regards,

K. Supraja (211FA04020)
K. Thanmai Ganga Bhavani(211FA04285)

iii
 ABSTRACT

Classifying multiple sounds in continuous audio streams is a complex and critical task,
especially in real-world scenarios where different audio sources such as background
conversations, music, environmental noise, and speech can occur simultaneously. Traditional
audio classification methods often struggle in such conditions due to the complexity and
variation in the sound patterns. Classifying multiple sounds in continuous audio streams is a
complex and critical task, especially in real-world scenarios where different audio sources
such as background conversations, music, environmental noise, and speech can occur
simultaneously. Traditional audio classification methods often struggle in such conditions due
to the complexity and variation in the sound patterns. This project presents a deep learning
based approach to effectively identify and categorize individual sound types within a mixed
audio recording. To process and analyze audio data, raw signals are first converted into
informative visual representations such as Mel-frequency Cepstral Coefficients (MFCCs) and
spectrograms, which record crucial time-frequency characteristics. Convolutional Neural
Networks (CNNs), which can automatically learn complex sound patterns and spatial
information, are then fed these representations. CNNs are excellent at extracting hierarchical
characteristics, which makes them ideal for jobs that require classifying continuous and
overlapping sounds. There are many real-world uses for this approach. In voice-controlled
smart homes, it can help detect unusual sounds like glass breaking or smoke alarms. In online
examination systems, it can be used to monitor background noise for potential violations such
as whispering or unauthorized conversations. Other applications include enhancing the
robustness of speech recognition systems, improving multimedia content indexing, and
supporting surveillance or security /systems by identifying critical sound events in real-time.

iv
 TABLE OF CONTENTS

1. Introduction 1
1.1 What is Audio Classification and what causes the need for it? 3
1.2 The consequences of inaccurate Audio Classification? 4
1.3 why is identifying multiple continuous sounds important? 5
1.4 Current Methodologies 5
1.5 Applications of Deep Learning in Audio Classification 6
2. Literature Survey 8
2.1 Literature review 9
2.2 Motivation 13
3. Proposed System 14
3.1 Input dataset 15
3.2 Data Pre-processing 15
3.2.1 Resampling 16
3.2.2 Noise Reduction/Silence Removal 17
3.2.3 Feature Extraction 17
3.3 Model Training 20
3.3.1 Splitting Data 20
3.4 Methodology of the system 20
3.5 Model Evaluation 22
3.5.1 Model Summary 23
3.6 Constraints 24
3.7 Cost and Sustainability Impact 26
4. Implementation 28
4.1 Environment Setup 29
4.2 Sample code 29
5. Experimentation and Result Analysis 33
5.1 Results 34
6. Conclusion 38
7. References 40

v
 LIST OF FIGURES

Figure 3.1Architecture of the proposed system 16

Figure 3.4.1 Methodology of the system 21

Figure .3.5.1 Actual vs Predicted Classes for Each Audio Segments 35

Figure 3.5.2 ROC Curve for Multi-Class Classification 37

vi
 LIST OF TABLES

Table: 3.5.1 Model Summary 23

1
CHAPTER-1
INTRODUCTION

2
 INTRODUCTION

1.1 What is Audio Classification and what causes the need for it?

Audio classification is the process of analyzing and categorizing different types of sounds
into specific classes such as speech, music, or everyday environmental sounds like dog
barking, horns, alarms, or children playing. This process typically begins by capturing raw
audio signals and converting them into useful data using techniques like Mel-Frequency
Cepstral Coefficients (MFCCs), spectrograms, or chroma features. Once features are
extracted, machine learning or deep learning models are used to identify and classify the
type of sound.

The need for audio classification has grown rapidly due to the increasing use of smart
technologies in our daily lives. Devices like smartphones, smart home systems,
surveillance cameras, and autonomous vehicles rely on sound to better understand and
respond to their surroundings. For example, a voice assistant must recognize whether
you're giving a command or playing music, while a surveillance system should be able to
detect suspicious sounds like a scream or a breaking window.

Audio classification is also crucial in areas like health monitoring (e.g., detecting coughs
in patients), wildlife conservation (monitoring animal calls), and public safety (identifying
alarms or sirens). Additionally, in entertainment and media, sound classification helps in
tagging content, enabling better content discovery and recommendation. In industrial
environments, it aids in machinery fault detection by identifying abnormal sounds. As
digital environments become more interactive, there's a growing demand for machines that
can "hear" and react accurately to various sounds in real time. This helps automate
processes, improve user experiences, and enable smarter, more responsive systems across
industries.

audio classification is becoming an essential component of modern intelligent systems.

Beyond its core role in identifying and categorizing sounds, it plays a vital part in
enhancing the functionality and responsiveness of technology across diverse sectors. As
artificial intelligence continues to evolve, audio classification contributes significantly to
human-computer interaction by enabling machines to interpret auditory cues just like
humans do. Furthermore, the integration of audio classification with other sensory data—
such as video or text—can lead to more robust multimodal systems capable of deeper

3
contextual understanding. This not only improves system accuracy but also opens the door
to innovations in areas like immersive gaming, virtual reality, real-time transcription
services, and assistive technologies for people with disabilities. As we move towards
smarter, more perceptive environments, the role of accurate, efficient, and real-time audio
classification will continue to grow in importance, driving forward the development of
truly intelligent systems.

1.2 The consequences of inaccurate Audio Classification?

Inaccurate audio classification can lead to significant consequences, especially in real-time
and safety-critical applications. When a system misinterprets or fails to detect a specific
sound, it may trigger false alarms or completely miss vital alerts. For instance, in
surveillance and security systems, if gunshots or distress calls are not correctly identified,
it could result in delayed responses from emergency services, potentially risking human
safety. In healthcare applications, misclassifying coughs or other symptoms could hinder
early diagnosis or monitoring of patients. In smart home environments, incorrect detection
of voice commands or environmental sounds may lead to device malfunction or poor user
experience.

Furthermore, in multimedia retrieval systems, wrongly labeled audio files can reduce the
efficiency and accuracy of content indexing and searching. Such inaccuracies also impact
training data for machine learning models, leading to further degradation of performance
over time. In business applications like customer service or call center monitoring,
misclassified audio inputs can affect analytics and customer satisfaction. Therefore,
maintaining high accuracy in audio classification is critical not only for operational
efficiency but also for user trust, system dependability, and real-world impact where sound
plays a vital role in communication and automation.

it is evident that inaccurate audio classification can have a profound ripple effect across
various sectors. In addition to the already highlighted domains, industries such as
automotive (e.g., driver-assist systems that rely on detecting sirens or honks), education
(e.g., transcription tools or lecture capture systems), and entertainment (e.g., automatic
tagging or sound mixing) also suffer when sound is not accurately interpreted.
Misclassification can lead to flawed decision-making, loss of valuable insights, and even

4
legal liabilities in cases where audio evidence plays a role. Furthermore, in environments
involving human-computer interaction, poor audio recognition diminishes accessibility for
individuals relying on voice-based systems, such as those with visual impairments.

1.3 Why is identifying multiple continuous sounds important?

Identifying multiple continuous sounds accurately is crucial because many real-world
applications involve environments with overlapping or sequential sounds that provide vital
information. In security systems, for example, distinguishing between various continuous
sounds like footsteps, breaking glass, or gunshots is essential for triggering appropriate
responses. Failure to identify these sounds accurately could result in delayed or incorrect
actions, potentially jeopardizing safety.

In healthcare, identifying continuous sounds like wheezing, coughing, or heartbeats is

critical for monitoring patient health, diagnosing conditions, and providing timely
interventions. In smart home environments, voice commands or ambient sounds (e.g., a
baby crying or a doorbell ringing) need to be recognized correctly to ensure seamless user
interaction and automation. Additionally, in industrial settings, accurate identification of
continuous machinery sounds can help in early detection of equipment failures, reducing
downtime and preventing accidents. Overall, identifying multiple continuous sounds
ensures that systems can react appropriately, ensuring efficiency, safety, and accuracy in
real-time decision-making.

1.4 Current Methodologies

Audio classification involves converting raw sound signals into meaningful features using
methods like MFCCs and spectrograms. These features are input into machine learning or
deep learning models to recognize different types of sounds. Convolutional Neural
Networks (CNNs) are widely used for analyzing spectrograms, capturing frequency-based
patterns. LSTM and other recurrent networks are effective for understanding temporal
sequences in audio, especially for sounds that evolve over time.For identifying
simultaneous or overlapping sounds, hybrid models combining CNN and LSTM are
popular. These models leverage CNN’s spatial and LSTM’s temporal strengths to detect
multiple sound events in a single clip.

5
While traditional methods like SVM and Random Forest are still in use, deep learning
models offer superior performance, especially with large datasets. Techniques like data
augmentation and ensemble learning further improve accuracy in noisy or real-world
environments.Despite progress, challenges remain in detecting low-volume or overlapping
sounds accurately, particularly in dynamic, noisy settings. Continued research in model
architectures and multimodal integration is helping to close these gaps.

1.5 Applications of Deep Learning in Audio Classification

Deep learning has revolutionized a wide range of audio classification applications by
enabling more intelligent and context-aware systems that can operate in real time and
under challenging conditions. One of the most impactful areas is voice-controlled
assistants, such as Google Assistant, Amazon Alexa, and Apple Siri. These systems rely
on deep neural networks, particularly LSTMs and transformer-based architectures, to
recognize and interpret voice commands with high accuracy—even in the presence of
background noise or accents. For example, advancements in Google's speech recognition
system have brought the word error rate down to levels comparable to human transcription
in some contexts.

In the field of smart surveillance, deep learning models are deployed to detect and classify
potentially dangerous sounds like gunshots, breaking glass, or distress calls. These systems
are widely used in urban public safety infrastructures. Technologies like ShotSpotter use
real-time audio classification to notify law enforcement of gunfire incidents with high
precision and low latency, helping authorities respond faster and more accurately.

Another critical area is environmental and noise monitoring, where urban planners and
local governments use deep learning to classify and track sound pollution sources.
Initiatives like the Sounds of New York City (SONYC) project use a network of sensors
and CNN-based models to analyze millions of audio clips, detecting sound types such as
jackhammers, honking, or sirens. These insights help authorities make informed decisions
about zoning, traffic regulation, and community well-being.

6
Deep learning also plays a significant role in the entertainment and music industry.
Music genre detection systems powered by CNNs and spectrogram analysis can identify
not only the genre but also instruments, tempo, and mood of a song. This automatic tagging
improves the effectiveness of recommendation engines on platforms like Spotify,
YouTube Music, and SoundCloud, personalizing user experience based on listening
history and preferences.

7
 CHAPTER-2
LITERATURE SURVEY

8
 2. LITERATURE SURVEY

2.1 Literature review

Indumathi et al. [1] engineered a deep learning-enabled system for classifying bird sounds
to tackle the difficulty of accurately recognizing bird species based on their vocalizations.
The research emphasizes the significance of employing convolutional neural networks
(CNNs) and long short-term memory (LSTM) models address the shortcomings of
conventional classification techniques. Their proposed approach leverages spectrogram
analysis along with transfer learning to boost classification accuracy. In this framework,
CNNs are used to extract features, while LSTMs handle the modeling of temporal
sequences, collectively enhancing the precision of bird species recognition.

Momynkulov et al. [2] Outlined a strategy driven by deep learning utilizing a CNN-RNN
model for detecting and classifying dangerous urban sounds. Their study employed the
ESC-50 dataset, selecting approximately 300 audio samples, including gunshots,
explosions, sirens, and cries. The results indicate the effectiveness of deep learning in real-
time security applications, enabling automated sound-based surveillance in public and
confined spaces. This method enhances law enforcement and emergency response by
efficiently identifying critical auditory events.

Garcia et al. [3] introduced a fall detection system based on machine learning, utilizing the
SAFE dataset, which includes 950 audio recordings, with 475 representing simulated fall
incidents. The study explored decision trees, Gaussian Naive Bayes, and deep learning
models incorporating spectrogram analysis for classification. Findings confirmed the
efficiency of audio-based fall detection, with deep learning models exhibiting the highest
performance. This approach aims to enhance elderly care by enabling real-time monitoring
of falls in residential and healthcare environments.

Aslam et al. [4]reviewed underwater sound classification Leveraging artificial intelligence

methods, analyzing datasets like ShipsEar, DeepShip, and the Watkins Marine Mammal
Sound Database. The study explored various methods, including SVM, KNN, CNNs, and
RNNs, highlighting their effectiveness in different scenarios. It found that deep learning
models excel in complex environments, while traditional ML performs better with smaller
datasets. The study predicts that AI-driven techniques will become the standard for marine

9
vessel detection, bioacoustics monitoring, and environmental assessments.

Brunese et al. [5] introduced a deep learning-based method for detecting heart disease by
analyzing cardiac sounds from the Classifying Heart Sounds Challenge dataset. Their ap
proach transforms audio signals into numerical features such as MFCCs and utilizes a deep
neural network for classification tasks. The results demonstrate that deep learning models
are highly effective in distinguishing between healthy individuals and those with heart
conditions, surpassing the performance of conventional techniques. The study also
suggests that mobile applications powered by deep learning could facilitate early diagnosis
and remote health monitoring, potentially minimizing the need for frequent hospital visits.

Kho et al. [6] developed a deep learning framework using the Mini VGG Net model for
COVID-19 detection through cough sound analysis. The study utilized datasets from the
University of Cambridge, Coswara Project, and NIH Malaysia. It incorporated cough
segmentation techniques and examined data augmentation effects, finding that
segmentation improved model performance, while augmentation had no significant
impact.

Pezzoli et al. [7]introduced a spherical harmonics-based MNMF framework for sound

source separation. Using simulated data sets with HOMs, the model isolates direct sound,
reducing reverberation effects. The approach showed superior performance over existing
techniques, with potential applications in virtual reality, spatial audio, and immersive
sound technologies.

Arafath et al. [8] developed a deep learning approach for breath sound detection using
speech recordings and thermal video data. Their method incorporated self-supervised learn
ing and CNN-BiLSTM models to enhance performance. The study suggests applications
in medical diagnostics, biometric authentication, and respiratory health monitoring.

Kim et al. [9] conducted milling experiments using varied machining parameters and
labeled chatter events using expert knowledge. They applied a CNN with an attention
block combining AlexNet outputs and cutting parameters. The model achieved 94.51%
accuracy in OOD testing, outperforming the baseline CNN model’s 88.66% accuracy.

10
The research carried out by Sophia et al. [10] explored how The use of audio elements like
sound effects and music enhances robotic storytelling in a range of genres. Four online
studies were conducted using stories from horror, detective, romance, and comedy genres.
The findings revealed that incorporating sound and music enhanced the enjoyment of
romantic stories and generally contributed to reduced fatigue across all genres.

Tuomas Virtanen et al. [11] used the TUT Urban Acoustic Scenes 2018 dataset to classify
short audio samples into ten acoustic scenes using a CNN-based baseline. The model
achieved 59.7% on the development set and up to 61% on evaluation, with reduced
accuracy on varied recording devices.

Yizhar Lavner et al. [12] conducted a study using a dataset comprising audio recordings
of infants aged 0 to 6 months in home settings. Their approach applied two machine learn
ing techniques for the automatic recognition of baby cries: a lightweight logistic regression
model and a convolutional neural network (CNN). The findings indicate that the CNN
significantly outperforms the logistic regression classifier in terms of detection accuracy.

Three publicly accessible ambient audio benchmark datasets are used in Sheryl Brahnam
et al.’s [13] study: (1) bird cries, (2) cat sounds, and (3) the ambient Sound Classification
(ESC 50) database. Five pre-trained convolutional neural networks (CNNs) are retrained
using ensembles of classifiers that em ploy four signal representations and six data
augmentation approaches. The findings demonstrate that the best ensembles perform better
than or in comparison to the best approaches described in the literature on several datasets,
including the ESC-5 dataset.

Piczak et al. [14] explored the application of convolutional neural networks (CNNs) for
the classification of brief audio segments featuring environmental sounds. Their method
employed a deep learning structure consisting of two convolutional layers, followed by
max-pooling operations and two fully connected layers. The model was trained using low-
level audio inputs, particularly segmented spectrograms enhanced with delta features. To
assess the effectiveness of their approach, the system was evaluated using three publicly
available datasets containing samples of environmental and urban audio.

Cakir et al. [15] proposed a deep learning-based method employing multi-label neural

11
networks to detect simultaneously occurring sound events in real-world acoustic scenes.
Their study utilized a diverse dataset consisting of over 1100 minutes of audio captured
from 10 different everyday environments, comprising 61 distinct sound event categories.
The approach incorporated log Mel-band energy features and a median filter ing technique
during post-processing. The model demonstrated an accuracy of 63.8%, achieving a 19%
improvement over conventional baseline systems.

Forrest Briggs et al. [16] introduced a technique to detect multiple bird species vocalizing
simultaneously by utilizing a multi-instance multi-label (MIML) classification framework.
The model was evaluated on 548 ten-second audio samples recorded in natural forest
settings. By applying a specialized segmentation method for feature extraction, their
approach achieved a 96.1% accuracy rate, effectively handling noisy conditions with
overlapping bird calls.

The tutorial by Mesaros et al. [17]explores sound event detection using deep learning
methods like CRNNs with log mel spectrogram features. It discusses datasets such as
Audio Set and URBAN-SED, and techniques like data augmentation, transfer learning,
and weak/strong labeling. Evaluation uses F-score, precision, recall, and metrics like PSDS
for accuracy.

Zaman et al. [18] conducted a review of deep learning models for audio classification.
Their work examines Convolutional Neural Networks (CNNs), Recurrent Neural
Networks (RNNs), autoencoders, transformers, and hybrid architectures. They discuss
audio datasets like ESC-50 and UrbanSound8k, and analyze the application of various
deep learning architectures to audio classification tasks.

In their work, Qamhan et al. [19] utilized the KSU-DB corpus, consisting of 3600
recordings from 3 environments and 4 recording devices, to build an acoustic source
identification system. This system, based on a hybrid CNN-LSTM model, was designed
to classify recording devices and environments. The study demonstrated that
voiced/unvoiced speech segments are effective for this classification task, with the system
reaching 98% accuracy for environment and 98.57% accuracy for microphone
classification.

12
Inoue et al. [20]applied anomaly detection methods that use classification confidence to
the DCASE 2020 Task 2 Challenge. Their systems ensemble two classification-based
detectors. They trained classifiers to classify sounds by ma chine type and ID, and to
classify transformed sounds by data-augmentation type. The dataset was The DCASE
2020 Task 2 Challenge

2.1 Motivation
The motivation for audio classification, particularly the identification of single and
multiple sounds within an audio clip, stems from the growing demand for intelligent
systems that can understand and interact with complex acoustic environments. In today’s
world, audio-based data is ubiquitous, ranging from urban soundscapes and natural
environments to speech and industrial settings. Accurately identifying sounds from such
diverse sources has significant implications in fields like surveillance, healthcare
monitoring, smart homes, autonomous vehicles, and human-computer interaction.

Traditional audio classification methods often struggle when multiple sounds occur
simultaneously, as overlapping acoustic signals can obscure or distort key features. This
limitation hinders their effectiveness in dynamic, real-world environments. Therefore,
there is a strong need for advanced approaches that can not only recognize isolated sounds
but also disentangle and correctly label multiple concurrent sound sources. By enabling
systems to perceive and differentiate between overlapping audio events, we move closer
to creating machines that can interpret sound as effectively as humans—understanding
context, reacting appropriately, and making intelligent decisions based on the auditory
scene. This capability is crucial for building responsive AI systems that operate reliably in
uncontrolled, noisy, and complex environments.

13
 CHAPTER-3
PROPOSED SYSTEM

14
 3. PROPOSED SYSTEM

In this study, we developed an advanced methodology for classifying multiple sounds

within a single audio clip using a pre-trained Convolutional Neural Network (CNN) model.
This model, specifically designed for sound classification tasks, processes input audio by
extracting key auditory features and then predicts the presence of various sound types. The
approach is robust enough to manage complex audio sequences that include overlapping
sounds, which makes it particularly effective for real-world scenarios such as
environmental sound recognition, speech detection, and general noise classification.

The entire process begins by segmenting the audio input into smaller portions. Each
segment is then fed into the CNN model, which has been pre-trained and stored in a file
named model.h5. This model operates in a multi-class classification mode, meaning it can
identify and distinguish between several sound classes simultaneously. For every segment,
the model outputs a probability distribution across the predefined sound classes.

The class with the highest probability is selected using the argmax function, effectively
assigning the most likely sound label to that segment. This predicted numerical label is
then decoded into its corresponding sound category using a LabelEncoder that maps
numeric classes back to their original names. The final output is a list of predicted sound
labels, one for each segment, providing a detailed breakdown of all the different types of
sounds detected throughout the audio clip. This method ensures accurate identification
even in audio environments with complex, overlapping auditory signals.

The model (fig-3.1) is structured around a Convolutional Neural Network

(CNN),specifically designed for the classification of 1D sequential data. It includes
multiple layers that capture hierarchical features from the input, followed by dense layers
that map these learned features to the corresponding output classes. Below, the specific
layers and processes of the model are outlined, along with the primary mathematical
expressions used to determine the outputs.

15
 Fig. 3.1. CNN Architecture

3.1 Input Dataset

The UrbanSound8K dataset is a well-known and widely used collection of audio clips
curated to support research in environmental sound classification and audio event
detection. It consists of 8,732 clips, each with a maximum duration of 4 seconds, although
some may be shorter de pending on the specific sound event recorded. The dataset occupies
approximately 5.27 GB of storage.
Each clip is labeled with one of 10 predefined sound categories, which include: Air
Conditioner, Car Horn, Children Playing, Dog Bark, Drilling, Engine Idling, Gunshot,
Jackhammer, Siren, and Street Music. These categories cover a broad range of everyday
urban and environmental sounds, spanning human and mechanical activities, animal
vocalizations, and ambient background noise. This diversity makes the UrbanSound8K
dataset highly representative of real-world audio scenes and valuable for developing and
evaluating models for audio-based applications.

3.2 Data Pre Processing

3.2.1 Resampling
Resampling is the process of altering the sampling rate of an audio signal, which refers to
the number of audio samples captured per second and is measured in Hertz (Hz). For

16
instance, audio recorded at 44,100 Hz (commonly used in CDs) contains 44,100 samples
per second. However, in many machine learning and deep learning applications, such a
high sampling rate may not be necessary and can lead to increased computational costs.
Resampling helps address this by reducing or standardizing the sampling rate (e.g., down
sampling to 16,000 Hz or 8,000 Hz), making the audio data more manageable and
consistent across datasets. It also ensures compatibility with pre-trained models or feature
extraction methods that expect input at a specific rate. Additionally, resampling can
enhance the generalization of models by eliminating redundant frequency information that
is not critical for the classification task.

3.2.2 Noise Reduction / Silence Removal

Noise Reduction / Silence Removal plays a vital role in improving the quality of input data
by removing non-informative parts of the audio. This includes background noise, silent
intervals, and other irrelevant sounds that do not contribute to the task of classification.
Noise, such as static hums, ambient chatter, or electronic interference, can obscure
important acoustic features and lead to misclassification. Silence segments, on the other
hand, unnecessarily inflate the length of input without adding value, slowing down both
training and inference processes. By applying techniques such as spectral gating,
threshold-based silence trimming, or noise profiling, these unhelpful elements are
minimized. As a result, the processed audio is cleaner, more focused, and better aligned
with the features that actually define the different sound classes. This enhances model
accuracy, reduces overfitting, and contributes to more efficient training and evaluation in
audio classification systems.

3.2.3 Feature Extraction

Mel-Frequency Cepstral Coefficients (MFCCs) are extensively used as feature

representations in audio classification and speech identification tasks. Rooted in the
nonlinear perception of sound by the human auditory system, MFCCs offer a concise
representation of an audio signal’s spectral characteristics. The process of extracting
MFCCs involves a series of signal processing stages, each designed to retain perceptually
significant information.

17
1. Pre-emphasis
Initially, the raw audio signal undergoes pre-emphasis filtering, which amplifies high-
frequency components to offset the natural attenuation introduced by the vocal tract or
environmental noise. The filtered output signal y[n] is defined by:

where β is a pre-emphasis coefficient typically chosen between 0.95 and 0.97.

2. Framing and Windowing

Following the pre-emphasis step, the audio signal is partitioned into overlapping time
frames, usually spanning 20 to 40 milliseconds. A Hamming window is subsequently
applied to each segment to suppress spectral distortion before executing the Fourier
transform.

where N indicates the length of the frame in terms of samples.

3. Fast Fourier Transform (FFT)

The Fast Fourier Transform (FFT) is applied to each windowed frame to convert it from
the time domain to the frequency domain:(FFT):

4. Power Spectrum and Mel Filter Bank

The power spectrum is passed through a bank of triangular filters that are spaced linearly
at low frequencies and logarithmically at higher frequencies, mimicking the human ear’s
frequency resolution. The mapping from frequency (Hz) to Mel scale is given by:

18
The output of each filter represents the energy in that Mel band.

5. Logarithmic Compression
To mimic the human ear’s logarithmic response to loudness, the energy within each Mel
frequency band is transformed using a logarithmic scale.

where Sm is the energy output of the mth Mel filter.

6. Discrete Cosine Transform (DCT)

To reduce redundancy among features and emphasize the most significant information,
DCT is applied to the log-scaled Mel spectrum. This transformation compacts the energy
into a smaller set of coefficients, producing the final MFCCs.

where M is Mel filters and L is the number of MFCCs retained (typically 12–13).

7. Post-Processing
In practical scenarios, temporal patterns can be captured by computing the first and
second-order derivatives, known as delta and delta-delta coefficients. In this research, 40
MFCCs were derived from each audio file, and the mean of each coefficient across all time
frames was computed to form a fixed-size feature vector. This vector was then utilized as
input to train a deep learning model for classifying urban sounds.

19
3.3 Model Training
The model was trained over 150 epochs using a batch size of 32. This batch size indicates
how many data samples are handled before the model adjusts its internal weights. The
training ran in verbose mode, meaning progress details such as loss and accuracy were
shown after every epoch.

3.3.1 Splitting Data

To prepare the dataset for model training and evaluation, the data was divided into two
subsets: training and testing. This was achieved using the train_test_split function
provided by the Scikit-learn library, a widely-used toolkit for machine learning in Python.
In this approach, 80% of the total dataset was allocated for training purposes, while the
remaining 20% was reserved for testing. This split ensures that the model is trained on a
substantial portion of the data to learn patterns and features effectively, while also being
evaluated on a separate, unseen subset to objectively measure its generalization capability.
To maintain reproducibility and ensure consistent results across multiple runs, the
parameter random_state=42 was specified. This sets a fixed seed for the random number
generator, guaranteeing that the same split is produced each time the code is executed.
Using a fixed random state is essential in experimental setups, especially when comparing
model performances or conducting ablation studies, as it removes variability introduced
by random sampling.
This train-test split strategy is a fundamental part of supervised learning workflows. It
allows for a realistic assessment of how well the model performs on new, unseen data,
which is critical for evaluating the model’s robustness and avoiding overfitting to the
training data.

3.4 Methodology of the system

The figure below illustrates the complete pipeline for audio classification using a
Convolutional Neural Network (CNN) model. The process begins with Data Collection,
which involves gathering a variety of audio recordings from relevant sources. These
recordings may include environmental sounds, speech, noise, or other acoustic events
depending on the objectives of the classification task.

20
 Fig. 3.4.1 Methodology of the system

Once collected, the audio data undergoes Pre-processing, a crucial step aimed at
improving the quality of the input. Pre-processing typically involves resampling to a
standard frequency, normalizing audio levels, removing silence or background noise, and
trimming unnecessary sections. The cleaned audio signals are then transformed into
Spectrograms, which are time-frequency visual representations of the audio. This visual
form helps capture both spectral and temporal patterns, making it more suitable for deep
learning models like CNNs.

Following this, Feature Extraction is performed. This step involves deriving meaningful
numerical representations from the spectrograms that can effectively characterize the
audio signal. One of the most commonly used features is the Mel Frequency Cepstral
Coefficients (MFCC), which compresses the audio signal into a lower-dimensional space
while preserving essential information relevant to human auditory perception. These
extracted features are saved in a structured format—typically as a CSV (Comma-
Separated Values) file for further processing.

21
The CSV file is then loaded into a Data Frame, a tabular structure used for organizing
and manipulating data efficiently using libraries such as Pandas in Python. At this stage,
the dataset is divided into training and testing subsets using Train & Test Split, generally
in an 80:20 ratio. This ensures that the CNN model is trained on a majority of the data
while being tested on unseen data to evaluate its generalization performance.

Once the data is split, it is passed into the CNN Model for training. The model learns to
identify patterns and correlations between the features and the corresponding sound labels.
After training, the CNN is used for Sound Prediction, where it classifies new, unseen
audio segments into predefined sound categories based on the learned features.

This end-to-end pipeline from raw audio input to final prediction demonstrates an effective
methodology for developing a robust audio classification system that can handle both
single and multiple sound events in real-world applications.

3.5 Model Evaluation

After completing the training phase, the performance of the CNN model is rigorously
evaluated using the test dataset, which consists of previously unseen audio samples. This
step is essential to assess the model’s generalization ability—that is, how well the model
performs on new data that it was not exposed to during training. Evaluating on the training
data alone would only measure how well the model memorized the data; thus, testing on a
separate dataset provides a more realistic estimate of its predictive power in real-world
applications.
The evaluation is conducted using the .evaluate() method, a built-in function provided by
deep learning frameworks such as Keras. This function computes various performance
metrics by comparing the model's predicted outputs against the true labels in the test
dataset. Among the metrics calculated, accuracy is one of the most fundamental and
interpretable. It is defined as the proportion of correctly predicted samples to the total
number of test samples.
In this study, the CNN model achieved an accuracy of 78% on the test set. This indicates
that 78% of the audio clips in the testing data were correctly classified into their respective
sound categories by the trained model. Such a result demonstrates a reasonably good

22
generalization capability, suggesting that the model has effectively learned meaningful
patterns from the training data and can apply them to new, unseen audio inputs with a fair
level of confidence.

3.5.1 Model Summary:

Table: 3.5.1 Model Summary

1. Input & Initial Convolution:

• conv1d: First 1D convolutional layer with 32 filters, producing output of shape

(None, 38, 32); 128 parameters.
• activation: Applies an activation function (likely ReLU).
• max_pooling1d: Reduces the dimensionality by half, to (None, 19, 32).
• dropout: Applies dropout regularization to prevent overfitting.

2. Second Convolution Block:

• conv1d_1: Another convolutional layer with 64 filters, resulting in (None, 17, 64);
6,208 parameters.
• activation_1, max_pooling1d_1, and dropout_1: Again apply non-linearity,
pooling, and regularization. Pooling reduces shape to (None, 8, 64).

23
3. Third Convolution Block:
• conv1d_2: Final convolution layer with 128 filters, resulting in (None, 6, 128);
24,704 parameters.
• activation_2, max_pooling1d_2, dropout_2: Shape reduced to (None, 3, 128) after
pooling.

4. Flattening and Dense Layers:

• flatten: Converts 3D tensor to 1D of shape (None, 384) for fully connected layers.
• dense: Fully connected layer with 100 units; 38,500 parameters.
• activation_3 and dropout_3: Apply non-linearity and regularization.
• dense_1: Final dense layer with 10 output classes; 1,010 parameters.
• activation_4: Applies softmax (or another final activation) to output class
probabilities.

3.6 Constraints
1.Continous Sound Events:
One of the most significant challenges is the presence of continuous sound events within
the same audio clip. When multiple sounds occur continously(e.g., a dog barking while a
vehicle passes by), it becomes difficult for models to isolate and correctly classify each
sound source. Traditional classification models typically assume a single dominant sound,
which limits their effectiveness in multi-label scenarios.

2.Environmental Noise and Distortions:

Real-world audio recordings often contain background noise, echo, or low-quality signals
due to poor recording conditions or environmental factors. These distortions can obscure
the primary sound features, reducing the model's ability to accurately distinguish and
classify relevant sounds.

3.Variability in Sound Characteristics:

Sounds of the same category can vary significantly based on factors like loudness, pitch,
duration, and recording device. For example, the sound of “footsteps” can differ widely

24
depending on the surface, speed, and footwear. Such variability introduces intra-class
variation that complicates model learning and generalization.

4.Imbalanced Data Distribution:

Datasets often have an uneven number of samples for different sound classes. Common
sounds (like speech or music) may dominate the dataset, while rare sounds (like sirens or
alarms) are underrepresented. This imbalance can lead to biased models that perform
poorly on underrepresented classes.

5.Temporal and Frequency Complexity:

Audio signals are inherently temporal and can exhibit complex frequency patterns.
Capturing both short-term and long-term dependencies is challenging, especially for short-
duration sounds that may be masked or missed entirely during feature extraction or
segmentation.

6.Annotation and Labeling Difficulties:

Manual labeling of multi-sound audio clips is labor-intensive and often subjective,
especially when sounds overlap or are partially audible. Incorrect or inconsistent labeling
can reduce the quality of training data, which in turn affects the model’s performance.

7.Model Limitations in Multi-label Classification:

Not all models are designed to handle multi-label outputs efficiently. Standard classifiers
predict only a single class per instance, whereas multi-sound clips require the model to
output multiple labels simultaneously, demanding more complex architectures and loss
functions.

8.Computational Resource Constraints:

Processing high-resolution audio or large datasets, especially when extracting complex
features (like spectrograms or MFCCs) and training deep models (like CNNs or RNNs),
can be computationally intensive. This may limit scalability or real-time application.

25
3.7 Cost and sustainability Impact
1.Computational Cost:
Training deep learning models such as Convolutional Neural Networks (CNNs) for audio
classification requires significant computational resources, particularly when dealing with
large datasets or high-resolution spectrograms. These costs are often associated with the
use of Graphics Processing Units (GPUs) or cloud-based services (e.g., AWS, Google
Cloud), which can incur substantial financial charges depending on training duration,
hardware configurations, and storage needs. Additionally, real-time or continuous audio
classification systems, such as those deployed in smart surveillance or monitoring systems,
demand ongoing inference capabilities, further increasing operational expenses.

2. Data Storage and Processing:

Audio data, especially when stored in high-quality formats or as extracted spectrogram
features, consumes considerable storage space. Continuous audio recording and
preprocessing can amplify these requirements, impacting both storage costs and data
management efforts. Moreover, preprocessing steps like noise reduction, resampling, and
feature extraction introduce additional computational overhead that must be accounted for
in both cost and energy consumption.

3. Sustainability and Environmental Impact:

The energy consumption of training deep learning models is non-trivial and contributes to
the overall carbon footprint of AI systems. Running large-scale training jobs on data
centers powered by non-renewable energy sources can lead to a significant environmental
burden. However, this impact can be mitigated by adopting energy-efficient practices, such
as using optimized models, applying transfer learning to reduce training time, and selecting
cloud providers that operate on green energy infrastructure. Inference at the edge (e.g., on-
device audio classification) can also improve sustainability by reducing reliance on
continuous cloud connectivity.

4.Long-Term Benefits and Cost Justification:

Despite the initial setup and training costs, audio classification systems offer long-term
cost efficiencies and environmental benefits when deployed strategically. For instance,
automated sound monitoring can reduce the need for manual supervision, lower human

26
resource expenditures, and enhance safety in environments such as factories, hospitals, and
wildlife monitoring zones. Moreover, early detection of critical audio events (e.g., alarms,
glass breaking, or screams) can lead to faster response times, potentially preventing
damage or saving lives jusifying the investment.

5.Scalability and Reusability:

Once trained, audio classification models can be reused and scaled across multiple
applications with minimal additional cost. Fine-tuning pre-trained models for specific
environments or sound categories enables faster deployment and reduces resource usage.
This approach supports sustainable AI development by promoting model longevity and
minimizing the need to retrain models from scratch.

27
 CHAPTER- 4
IMPLEMENTATION

28
 4.IMPLEMENTATION

4.1 Envirnomental setup:

To run sucessfully the script effectively, a proper environment setup is essential. The script
is compatible with Python version 3.8 or higher. It requires several key libraries, including
NumPy, Librosa, TensorFlow, and Scikit-learn, which should be installed using pip.
Utilizing a virtual environment is recommended to manage dependencies cleanly and
avoid version conflicts with other projects.In terms of hardware, a system with a minimum
of 8 GB RAM is necessary, though 16 GB is ideal for handling large audio datasets. While
the script can be executed on a CPU, using a GPU is highly beneficial for accelerating the
training process, especially when working with deep learning models.

The dataset structure is critical for successful execution. Audio files must be organized
into subdirectories within a main dataset folder. Each subdirectory should represent a
unique class label and contain audio samples specific to that class. For example, a folder
named “dog” could include barking sounds, while another labeled “ar” might contain
engine or horn sounds. This format enables the script to automatically read and label the
data correctly, facilitating efficient feature extraction, model training, and classification.
Proper setup ensures accurate audio classification and smooth workflow execution.

4.2 Sample Code

Importing Required Libraries
import os
import librosa
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder
from tensorflow.keras import models, layers

29
4.2 Feature Extraction from Audio Files
def extract_features(file_path):
audio, sample_rate = librosa.load(file_path, res_type='kaiser_fast')
mfccs = librosa.feature.mfcc(y=audio, sr=sample_rate, n_mfcc=40)
return np.mean(mfccs.T, axis=0)

4.3 Loading Dataset and Extracting Features

metadata = pd.read_csv('UrbanSound8K.csv')
features = []
labels = []

# Iterate and extract features

for index, row in metadata.iterrows():
file_path = os.path.join('audio', 'fold'+str(row["fold"]), row["slice_file_name"])
label = row["class"]
try:
data = extract_features(file_path)
features.append(data)
labels.append(label)
except Exception as e:
print(f"Error processing {file_path}: {e}")

4.4 Preparing Data for Training

X = np.array(features)
y = np.array(labels)
le = LabelEncoder()
yy = le.fit_transform(y)
y_encoded = tf.keras.utils.to_categorical(yy)

# Split data
X_train, X_test, y_train, y_test = train_test_split(X, y_encoded, test_size=0.2)

30
4.5 Building the Neural Network Model
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv1D, MaxPooling1D, Flatten, Dense, Dropout,
Activation
input_shape = (40, 1) # 1D data format
model = Sequential()

model.add(Conv1D(32, kernel_size=3, input_shape=input_shape))

model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Dropout(0.5))

model.add(Conv1D(64, kernel_size=3))
model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Dropout(0.5))

model.add(Conv1D(128, kernel_size=3))
model.add(Activation('relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Dropout(0.5))

model.add(Flatten())
model.add(Dense(100))
model.add(Activation('relu'))
model.add(Dropout(0.5))

model.add(Dense(num_labels))
model.add(Activation('sigmoid'))

model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

31
4.6 Compiling and Training the Model
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
model.fit(X_train, y_train, batch_size=32, epochs=50, validation_data=(X_test, y_test))

32
 CHAPTER- 5
RESULTS AND ANALYSIS

33
 5.RESULTS AND ANALYSIS

5.1 Results:

The process begins by taking the audio file located at the specified file path,
"/content/drive/MyDrive/continue/New folder/playing-dog.wav", which is stored in
Google Drive. This audio file is then fed into the trained model for analysis. After
processing the audio, the model identifies and classifies the sounds present in the file. As
shown in the output, the model successfully detects two distinct sound events: "children
playing" and "dog bark". This indicates that the model is capable of segmenting the audio
and accurately predicting multiple overlapping sound classes, demonstrating its
effectiveness in real-world sound classification tasks.

A. Model Evaluation and Test Accuracy:

After training the model, the performance is evaluated on the test dataset to assess its

generalization capabilities. The evaluation is carried out using the (evaluate) method, which

computes the accuracy of 78% on the test set.

B. Visualization of Actual vs Predicted Classes:

This demonstrates how to visualize and compare the predicted classes with the actual classes

for each audio segment. This visualization helps to assess the performance of a classification

model and identify discrepancies between predicted and true values.

The output graph for audio classification is generated by first dividing the input audio file

into overlapping segments, each five seconds long. For each segment, Mel-Frequency

Cepstral Coefficients (MFCCs) are extracted to capture key sound features, which are then

fed into a pre-trained deep learning model to predict the class of sound present. These

34
predicted class labels are collected in sequence, corresponding to each time segment of the

audio, forming the basis for visual comparison against actual labels.

The graph plots the predicted classes alongside the actual classes across all segments, with

each point representing a segment's classification result. This visual comparison helps

quickly identify where the model’s predictions match or differ from the expected outcome.

Matching lines indicate correct classification, while deviations reveal potential errors. The

plot, combined with the printed summary of actual vs. predicted values, provides a clear and

concise way to evaluate the model’s performance and analyze classification trends across

the duration of the audio file.

Fig.5.1.1 Actual vs Predicted Classes for Each Audio Segments

The visual output, as shown in figure, helps to quickly assess how well the model’s

predictions align with the true class labels for each audio segment.

35
C. ROC Curve for Multi-Class Classification:

In this section, the process of evaluating the performance of a multi-class classification

model is explained using the Receiver Operating Characteristic (ROC) curve and the Area

Under the Curve (AUC) score for each class.

The audio classification pipeline begins by dividing a continuous audio file into

overlapping segments to ensure that even brief or transitional sounds are captured. Each

segment is then processed using MFCC (Mel-Frequency Cepstral Coefficients) feature

extraction, which transforms raw audio into a compact numerical representation that

captures the spectral properties of sound. These MFCC features are reshaped and fed into

a trained deep learning model that predicts the probability of each sound class (e.g., dog

bark, siren, car horn). The predicted label for each segment is determined by selecting the

class with the highest probability. This process results in a list of predicted sound classes

for the entire audio file, which can be used for performance evaluation and visualization.

To assess the model’s effectiveness across different sound categories, a multi-class ROC

(Receiver Operating Characteristic) curve is plotted. This involves comparing the model’s

predicted probabilities against the true class labels in a one-vs-rest format by converting

the true labels into binary form for each class. The ROC curve for each class shows how

well the model distinguishes that specific sound from all others by plotting the true positive

rate against the false positive rate. The area under each curve (AUC) quantifies this

performance—where an AUC closer to 1.0 indicates excellent discrimination ability. This

visualization allows for a clear comparison of model performance across all sound

categories and helps identify which classes are classified reliably and which may need

further model improvement or more training data.

36
 Fig.5.1.2. ROC Curve for Multi-Class Classification

The graphical representation in Figure above provides insight into the model’s

performance across various sound categories. The Area Under the Curve (AUC) scores

reflect the model’s ability to differentiate each class from the others, with higher AUC

values signifying stronger classification accuracy.

37
CHAPTER-6

CONCLUSION

38
 CONCLUSION

We presented a Convolutional Neural Network (CNN) model in this paper that was
especially created to classify audio signals into several groups, including sounds from air
conditioners, automobile horns, kids playing, and other background noises. The design of
the model includes a number of layers, including dropout layers to avoid overfitting, max
pooling layers to minimize dimensionality, activation layers to introduce non-linearity, and
convolutional layers for feature extraction. Together, these layers gradually create a com
prehensive understanding of the audio input, allowing the model to identify ever-more
intricate patterns as it progresses through the layers. The approach eliminates the
requirement for manual feature engineering by automatically extracting relevant
characteristics from the raw audio input. Through the use of convolutional filters, it detects
important patterns that define various sound components. The pooling layers improve the
generalization of the model and reduce the computing cost. In order to increase the model’s
ability to function well on unseen data, dropout layers help to reduce overfitting. The f inal
classification result is then produced by combining the learnt features by the fully linked
layers at the model’s output. Performance metrics like accuracy and the ROC curve were
employed to evaluate the model’s efficacy; the Area Under the Curve (AUC) was utilized
to show how well the model was able to differentiate between different audio classes. By
correctly recognizing and categorizing audio signals using the learnt characteristics, the
model showed strong performance. The architecture is well-suited for real-world audio
categoriza tion applications because of its balanced design, which guar antees both
powerful performance and efficient processing. In conclusion, this CNN-based model
demonstrates its ability to recognize a variety of audio signals and provides a dependable
method for audio classification. Future research might focus on improving the model’s
generalization by experimenting with various hyperparameters, deepening the network, or
using data augmentation techniques. Exciting opportunities for further development and
implementation are presented by the model’s practical applications in domains such as
audio monitoring, smart home systems, and surveillance.

39
 REFERENCES

[1] I. Indumathi, S. Deepa, and S. S. Sridhar, “Deep learning-based bird sound

classification using CNN and LSTM with spectrogram and transfer learning,” Artificial
Intelligence in Signal Processing, pp. 233 597–603, 2024.
[2] Momynkulov, M. Kim, and H. Lee, “Deep learning-based detection and classification
of dangerous urban sounds using CNN-RNN models,” Expert Systems with Applications,
2023.
[3] J. Garcia, M. Lopez, R. Singh, and A. Patel, “Audio-based fall detec tion using machine
learning and spectrogram analysis on the SAFE dataset,” Multimedia Tools and
Applications,Volume 2, Issue 2, Au gust–December 2024, 100085.
[4] M. Aslam, R. Khan, A. Tariq, and L. Ahmed, “A review on under water sound
classification using machine learning and deep learning techniques,” Multimedia Tools
and Applications, Volume 255, Part A, 1 December 2024, 124498.
[5] L. Brunese, F. Mercaldo, A. Reginelli, and A. Santone, “Deep learning approach for
heart disease detection using cardiac sounds and MFCC features,” Multimedia Tools and
Applications, pp. 176 2202–2211,2020.
[6] J. Kho, L. Tan, M. Rahman, and S. Lim, “COVID-19 detection using cough sound
analysis with Mini VGGNet and segmentation techniques,” Multimedia Tools and
Applications,Volume 9, 2024, 100129.
[7] S. Pezzoli, M. Tylka, H. Kim, and V. Pulkki, “Spherical harmonics based MNMF
framework for sound source separation using higher-order microphone signals,”
Multimedia Tools and Applications, pp.218-109 888, 2024.
[8] M. Arafath, R. Mehmood, S. Iqbal, and N. Ahmed, “Deep learn ing approach for breath
sound detection using speech recordings and thermal video data,” IEEE Transactions on
Consumer Electronics, 3085–3093,2024.
[9] D. Kim, J. Lee, S. Park, and H. Choi, “Attention-based CNN frame work for chatter
detection using multimodal machining data,” IEEE Access,1386–1397, 2024.
[10] S. Sophia, J. Smith, L. Zhang, and K. M¨ uller, “Influence of sound effects and
background music in robotic storytelling across genres,” IEEE Access, pp., 2023.
[11] T. Virtanen, A. Mesaros, T. Heittola, and M. D. Plumbley, “Acoustic scene
classification using the TUT Urban Acoustic Scenes 2018 dataset,” IEEE Access, 2018.

40
[12] Y. Lavner, R. Rosenhouse, and A. Cohen, “Baby cry detection using CNN and
logistic regression in domestic audio recordings,” IEEE Access, pp.(202) 452-1535,2012.
[13] S. Brahnam, L. Jain, and T. Lee, “Environmental sound classification using CNN
ensembles and data augmentation,” IEEE Access, pp.(202) 452-1535 , 2012.
[14] J. Piczak, “Environmental sound classification with convolutional neural networks,”
IEEE Access, pp.254 207 621234,2010.
[15] E. Cakir, T. Heittola, H. Huttunen, and T. Virtanen, “Polyphonic sound event
detection using multi label deep neural networks,” in IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 2986–2990.
[16] F. Briggs, B. Lakshminarayanan, L. Neal, X. Z. Fern, R. Raich, S. J. K. Hadley, A. S.
Hadley, and M. G. Betts, “Acoustic classification of multiple simultaneous bird species:
A multi-instance multi-label approach,” J. Acoust. Soc. Am., vol. 131, no. 6, pp. 4640–
4650, 2012.
[17] A. Mesaros, T. Heittola, T. Virtanen, and M. D. Plumbley, “Sound Event Detection:
A Tutorial,” IEEE Signal Processing Magazine, vol. 38, no. 5, pp. 67–83, Sep. 2021, doi:
10.1109/MSP.2021.3090678.
[18] K. Zaman, M. Sah, C. Direkoglu, and M. Unoki, “A Survey of Audio Classification
Using Deep Learning,” IEEE Access, vol. 11, 2023.
[19] M. A. Qamhan, H. Altaheri, A. H. Meftah, G. Muhammad, and Y. A. Alotaibi,
“Digital Audio Forensics: Microphone and Environment Classification Using Deep
Learning,” IEEE Access, vol. 9, 2021, pp. 62719-62738.
[20] T. Inoue, P. Vinayavekhin, S. Morikuni, S. Wang, T. H. Trong, D. Wood, M.
Tatsubori, and R. Tachibana, “Detection of Anomalous Sounds for Machine Condition
Monitoring Using Classification Confidence,” Detection and Classification of Acoustic
Scenes and Events 2020, Tokyo, Japan, 2020.

41