INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 318
Sign Language Recognition Using Deep Learning: Advancements
and Challenges
David Bamidele Adewole, Ademola Adesugba, Olutola Agbelusi & Olukemi Victoria Olatunde
Department of Software Engineering, Federal University of Technology, Akure, Nigeria
DOI : https://doi.org/10.51583/IJLTEMAS.2024.131230
Received: 01 January 2025; Accepted: 06 January 2025; Published: 21 January 2025
Abstract: Sign language recognition (SLR) has arisen as a major area of research in recent years, attempting to bridge the
communication gap between the deaf and hard-of-hearing community and the hearing world. This research study addresses the
construction and implementation of a manual alphabet recognition system utilising deep learning techniques, notably
convolutional neural networks (CNNs). The work focuses on establishing an efficient and accurate system for converting
Nigerian Sign Language manual alphabets into text. By integrating computer vision and machine learning methods, the proposed
system seeks to overcome the communication gap between deaf and hearing individuals. The paper explains the technique
adopted, including data collection, preprocessing, model architecture, and deployment using web-based tools. The system
achieves a 95% success rate in recognizing static hand motions, proving its potential for real-world applications. However, issues
in identifying dynamic motions and generalizing across varied user populations are observed. The report finishes with
recommendations for future research, emphasizing the need for combining temporal analysis and expanding the system's
capabilities to word and phrase recognition.
Keywords: Deep Learning, Sign Language, Deaf, CNN
I. Introduction
Sign language serves as an essential means of communication for millions of deaf and hard-of-hearing individuals globally. The
World Health Organisation estimates that over 466 million individuals globally experience significant hearing loss, a number
projected to rise to 900 million by 2050 (WHO, 2024). In Nigeria, it is estimated that over 8.5 million individuals are deaf or hard
of hearing. Sign language constitutes the primary mode of communication for numerous individuals, enabling self-expression and
interaction with others (Eleweke, 2002; Asonye et al., 2018; Asonye et al., 2020). Nonetheless, the communication barrier
between sign language users and non-sign language users persists as a significant impediment. This barrier may result in social
isolation, restricted access to educational and employment opportunities, and challenges in obtaining necessary assistance. To
address this difficulty, academics and engineers have been exploring various approaches to develop sign language recognition
systems that can bridge this communication gap.
Sign language is a complicated visual-gestural language that uses hand forms, gestures, facial expressions, and body postures to
convey message. Sign languages are not universal; various countries and regions have developed their own distinct forms over
time (Simon, 1982; Karbasi et al., 2015; Cohen, 2020). This study focusses on the Nigerian Sign Language (NSL), a variant of
American Sign Language (ASL) that has been tailored to fit the cultural context of Nigeria. Manual alphabets, often known as
fingerspelling, are a fundamental component of sign languages. They are used to spell out words, names, or concepts that do not
have specific signs. In NSL, as in many other sign languages, the manual alphabet consists of 26 hand forms corresponding to the
letters of the English alphabet. With the development of artificial intelligence and machine learning technology, there has been
rising interest in developing automatic sign language recognition systems. These systems try to interpret sign language gestures
and translate them into text or voice, improving communication between deaf and hearing individuals (Gordon et al., 2005;
Joudaki et al., 2014; Dabwan, 2024).
Despite the improvements in technology, accurate and real-time sign language identification remains a demanding effort. This is
due to several factors:
a. Complexity of sign language: Sign languages are not merely visual representations of spoken languages but have their
own grammar, syntax, and lexicon.
b. Variability in gestures: Different signers may perform the same sign with small changes in hand shape, movement, or
posture.
c. Dynamic nature of signs: Many signs involve motion, making it problematic for static image-based recognition systems
to capture their full meaning.
d. Environmental factors: Lighting conditions, background clutter, and occlusions might decrease the accuracy of vision-
based recognition systems.
e. Limited datasets: There is a scarcity of substantial, diverse datasets for training machine learning models, particularly for
sign languages other than ASL.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 319
This research aims at developing a deep learning-based system for recognizing and translating Nigerian Sign Language manual
alphabets into text, to evaluate the performance of the system in terms of accuracy, speed, and robustness, to identify challenges
and limitations in the current approach and propose potential solutions and to contribute to the broader field of sign language
recognition research by sharing insights and methodologies.
II. Literature Review
The area of sign language identification has undergone great improvements in recent years, pushed by progress in computer
vision, machine learning, and deep learning technologies. This section gives an overview of relevant research in the issue,
focusing on approaches to manual alphabet recognition and bigger sign language translation systems.
Traditional Approaches to Sign Language Recognition: Early attempts at sign language recognition centred on hardware-
based solutions such as sensor gloves. Swee et al. (2007) created a "Wireless Data Gloves Malay Sign Language Recognition
System" employing gloves integrated with accelerometers and flexure sensors. While this strategy produced outstanding accuracy
for a small set of signs, it was impractical for wider implementation due to the cost and hassle of specialized technology. Color-
coded gloves were another approach investigated by researchers. Greenberg et al. (2015) described a method for detecting ASL
signs using inexpensive cotton gloves with distinct colours marking the base and fingers. This approach obtained 74% accuracy
for isolated sign recognition and 60% for continuous recognition. While more user-friendly than sensor gloves, this technology
still needed users to wear specialized equipment.
Vision-Based Approaches: As computer vision technologies improved, researchers began devising bare-hand approaches that
did not require specialized equipment. Paulraj et al. (2010) showed a phoneme-based sign language recognition system
leveraging skin color segmentation. Their method got a maximum classification accuracy of 92.85% for nine English phoneme
gestures. More recent studies have utilised deep learning techniques, particularly Convolutional Neural Networks (CNNs), for
sign language detection. Deshpande et al. (2023) demonstrated a real-time sign language recognition system leveraging CNNs to
capture and recognize sign language gestures. Their technique requires two key steps: gesture capture and CNN-based processing
to translate gestures into text and speech.
Manual Alphabet Recognition: Several studies have concentrated mostly on manual alphabet recognition. Oguntimilehin and
Balogun (2024) created an American Sign Language (ASL) fingerspelling translator employing a CNN with a pre-trained
GoogLeNet architecture. Their method provided solid categorisation results with new users, exhibiting effective performance
with less data. Shin et al. (2021) suggested a system for recognising ASL alphabets by extracting features from hand posture
predictions using MediaPipe. Their technique obtained 99.39% accuracy on the Massey dataset, 87.60% on the ASL Alphabet
dataset, and 98.45% on the Finger Spelling A dataset.
Deep Learning Approaches: Deep learning has emerged as a prominent tool for sign language recognition due to its capacity to
automatically uncover essential features from raw data. Zhang and Jiang et al. (2024) examined boosting ASL recognition with
deep learning models with transfer learning, examining architectures such as VGG16, ResNet50, MobileNetV2, and InceptionV3.
Their analysis suggested that InceptionV3 attained the best accuracy of 96%. Pathan et al. (2023) created a multi-headed CNN for
ASL recognition, employing image data and hand landmarks to increase detection accuracy. Their model attained a high-test
accuracy of 98.98% for recognizing static hand movements.
Real-Time Recognition Systems: Real-time recognition is crucial for practical applications of sign language translation systems.
Alaftekin et al. (2024) constructed a high-speed, accurate real-time hand gesture identification system for Turkish Sign Language
employing the YOLOv4-CSP algorithm. Their model scored 98.95% precision, 98.15% recall, 98.55 F1 score, and 99.49% mAP
in 9.8ms, displaying exceptional performance in both speed and accuracy.
Despite these developments, significant obstacles remain in the field of sign language recognition:
a. Dynamic gesture recognition: Most systems excel at recognizing static gestures but struggle with dynamic indications
that entail motion over time.
b. Generalization: Many systems perform well on specialised datasets but may not generalize successfully to varied user
populations or real-world settings.
c. Continuous sign language recognition: Recognizing individual signs or letters is different from reading continuous sign
language sentences, which entails understanding syntax and context.
d. Limited datasets: There is a scarcity of substantial, diverse datasets for many sign languages, particularly for languages
other than ASL.
e. Real-time performance: Balancing accuracy with speed remains a problem, especially for deployment on mobile or edge
devices.
This assessment of related works emphasises the progress made in sign language identification while also indicating areas for
further research and development. The current work intends to build upon these gains while addressing some of the noted
problems, specifically in the context of Nigerian Sign Language.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 320
III. Methodology
This section discusses the methodological approach utilised in constructing the manual alphabet recognition system for Nigerian
Sign Language. The methodology encompasses data collection, preprocessing, model architecture, training, and deployment.
Data Collection and Preprocessing
Dataset Creation: To train the deep learning model, a large dataset of Nigerian Sign Language manual letter movements was
produced. The dataset collecting technique involves the following steps:
i. Participant recruitment: A heterogeneous group of 20 native NSL signers was recruited to conduct the manual alphabet
movements.
ii. Image capture: High-resolution photographs were captured using a digital camera under varied lighting settings and
backgrounds to ensure diversity in the dataset.
iii. Gesture changes: Participants were instructed to create each letter multiple times with slight alterations in hand position
and orientation to improve the model's resilience.
iv. Dynamic gesture capture: For letters involving motion (such as J and Z), numerous images were taken to show different
stages of the gesture.
The first dataset included of roughly 50,000 photographs. After comprehensive review and deletion of obscure or illegible
signage, the final dataset was refined to 13,000 photographs, with 500 images each letter (A-Z).
Data Preprocessing: To prepare the dataset for training, the following preprocessing techniques were applied:
i. Image resizing: All photographs were shrunk to 128x128 pixels to preserve constant input size for the neural network.
ii. Normalization: Pixel values were normalized to the range [0, 1] by dividing by 255.
iii. Data augmentation: To improve the dataset's diversity and prevent overfitting, data augmentation procedures were
employed, including random rotations (±15 degrees), horizontal flips, and slight variations in brightness and contrast.
iv. Splitting: The dataset was partitioned into training (80%), validation (10%), and test (10%) sets.
Model Architecture
The manual alphabet recognition system is built on a Convolutional Neural Network (CNN) architecture (Fig. 1), which has
proven exceptional performance in picture classification applications. The network architecture is as follows:
Fig.1: CNN Architecture
The incorporation of numerous convolutional layers allows the network to learn hierarchical features, from low-level edge
detectors to high-level shape recognizers. The max pooling layers help reduce spatial dimensions and computational complexity.
Dropout layers are used to prevent overfitting.
a. Model Training the model was trained using the following parameters:
i. Optimizer: Adam (learning rate = 0.001)
ii. Loss function: Categorical cross-entropy
iii. Batch size: 32
iv. Epochs: 50 (with early halting dependant on validation loss)
Training was performed on a GPU-accelerated workstation to reduce calculation time. The model's performance was tracked
using accuracy and loss measures on both the training and validation sets.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 321
b. Hand Detection and Region of Interest Extraction For real-time recognition, the system employs the MediaPipe Hands library
to recognise and track hands in the input video stream. The hand detecting procedure requires two stages:
i. Palm detection: A single-shot detector model locates palms in the image.
ii. Hand landmark model: Once the palm is recognised, a second model predicts 21 3D hand landmarks.
The identified hand region is subsequently retrieved as the area of interest for categorisation by the trained CNN model.
System Deployment
The manual alphabet recognition system was deployed as a web application using the Streamlit framework. This choice allows
for easy access across different platforms and devices. The deployment process involved the following steps:
i. Model serialization: The trained CNN model was saved in a format compatible with web deployment.
ii. Web interface development: A user-friendly interface was created using Streamlit, allowing users to interact with the
system through their device's camera.
iii. Real-time processing: The application captures live video frames, performs hand detection and region of interest
extraction, and feeds the processed images to the CNN model for classification.
iv. Result display: The recognized letter is displayed in real-time on the web interface.
System Evaluation
To test the performance of the manual alphabet recognition system, the following measures were used:
i. Accuracy: The fraction of successfully categorised gestures in the test set.
ii. Confusion matrix: A detailed analysis of the model's performance for each letter.
iii. Precision, Recall, and F1-score: These measures provide a more nuanced perspective of the model's performance,
especially for imbalanced classes.
iv. Inference time: The time taken to process a single frame and produce a categorisation result.
By utilising this complete technique, the project intends to establish a robust and accurate system for detecting Nigerian Sign
Language manual alphabets, while also providing insights into the obstacles and prospects in this field.
IV. Results and Discussion
In this section, the performance results of the manual alphabet recognition system are shown and discussed in relation to the
research objectives and relevant literature.
Model Performance
Training and Validation Results: The CNN model was trained for 50 epochs, with early stopping employed to prevent overfitting.
Fig. 2 displays the training and validation accuracy over the course of training
Fig. 2: Training and Validation Accuracy
The model attained a final training accuracy of 99.2% and a validation accuracy of 87.8%. The high validation accuracy shows
that the model generalizes effectively to unknown data.
Fig. 3 depicts the training and validation loss across the training period.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 322
Fig. 3: Training and Validation Loss
The decreasing loss curves indicate that the model successfully learned to classify the manual alphabet gestures. The close
alignment between training and validation loss suggests that overfitting was effectively mitigated.
Test Set Performance:
On the held-out test set, the model attained an overall accuracy of 95%. Table 1 displays the precision, recall, and F1-score for
each letter.
Table 1: Precision, Recall, and F1-score for each letter
The model performed exceptionally well for most letters, with F1-scores above 0.95. However, some letters, particularly those
with similar hand shapes or those involving motion (e.g., J and Z), showed slightly lower performance.
Confusion Matrix
The confusion matrix reveals that most misclassifications occur between visually similar letters. For example, there is some
confusion between 'M' and 'N', and between 'S' and 'T'. This shows that the model might benefit from additional training data or
feature engineering to better distinguish between these similar hand shapes. Some notable observations from the confusion matrix
include:
a. The letters 'A', 'B', 'C', 'L', 'O', 'V', and 'Y' achieved perfect classification with no misclassifications.
b. 'J' and 'Z', which involve motion, showed lower accuracy compared to static gestures. This highlights the limitation of the
current model in capturing dynamic movements.
c. There was minor confusion between 'D' and 'F', likely due to the similarity in finger positioning.
d. 'I' and 'J' showed some mutual misclassification, possibly due to their similar starting positions.
e. 'K' and 'V' had a few instances of misclassification, which could be attributed to the similar extended finger positions.
To address these misclassifications, potential improvements could include:
a. Increasing the diversity of training data for commonly confused letter pairs.
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 323
b. Implementing data augmentation techniques to create more variations of challenging gestures.
c. Exploring advanced architectures or ensemble methods to capture more nuanced features distinguishing similar hand shapes.
Real-Time Performance
The system's real-time performance was evaluated based on its ability to process and classify gestures from live video input. The
average processing time per frame was measured at 0.05 seconds, allowing for a smooth experience of about 20 frames per
second. This performance is suitable for real-time applications, providing users with near-instantaneous feedback on their signed
gestures. However, it's worth noting that performance may vary depending on the hardware specifications of the user's device. On
less powerful systems, there might be a slight lag in recognition, which could impact the user experience.
User Experience Evaluation
To assess the system's usability and effectiveness in real-world scenarios, a small-scale user study was conducted with 10
participants, including both native sign language users and beginners. Participants were asked to perform a series of manual
alphabet gestures and provide feedback on the system's accuracy, responsiveness, and overall user experience. Key findings from
the user study include:
a. Native signers reported an average satisfaction score of 4.2 out of 5, praising the system's accuracy for most static gestures.
b. Beginners found the system helpful as a learning tool, with an average satisfaction score of 4.5 out of 5.
c. Both groups noted difficulties with dynamic gestures like 'J' and 'Z', confirming the quantitative results.
d. Users appreciated the real-time feedback, which allowed them to adjust their hand positions for better recognition.
e. Some users with darker skin tones reported occasional difficulties in hand detection under low lighting conditions, suggesting
a need for further optimization of the hand detection algorithm.
These user insights provide valuable direction for future improvements, particularly in enhancing the system's robustness across
different user demographics and environmental conditions.
Limitations and Challenges
While the manual alphabet recognition system produced promising results, numerous limitations and challenges were recognised
during the development and testing phases:
a. Dynamic Gesture Recognition: The current model struggles with gestures that entail motion, such as 'J' and 'Z'. This issue
derives from the use of static image classification, which doesn't capture temporal information.
b. Lighting and Background Sensitivity: The performance of the system can be impacted by varied lighting conditions and
complicated backgrounds, potentially lowering hand detection accuracy.
c. User Variability: Hand sizes, skin tones, and individual signing techniques might vary widely among users, providing issues
for the model's generalization capabilities.
d. Limited Vocabulary: The current method is restricted to identifying individual letters of the manual alphabet and does not
extend to whole words or sentences in sign language.
e. Computational Requirements: While the system operates well on typical desktop computers, it may experience performance
challenges on less capable machines, restricting its accessibility.
f. Occlusion Handling: The system may struggle when parts of the hand are obscured or when numerous hands are present in
the frame.
Addressing these restrictions will be vital for enhancing the system's robustness and expanding its practical applications. The next
chapter will examine various remedies and future directions to overcome these issues.
V. Conclusion
This work created and implemented a manual alphabet recognition system for Nigerian Sign Language employing deep learning
techniques, mainly convolutional neural networks (CNNs). The system demonstrates outstanding performance in recognizing
static hand motions representing letters of the manual alphabet, obtaining an overall accuracy of 95% on the test dataset. By
leveraging computer vision and machine learning techniques, the recommended approach makes a huge step towards bridging the
communication gap between deaf and hearing individuals in Nigeria. The methods adopted in this research, including meticulous
data collecting, preprocessing, model architecture design, and deployment using web-based tools, provides a solid platform for
future work in sign language recognition. The usage of MediaPipe for hand recognition and landmark identification, paired with a
bespoke CNN for classification, was effective in capturing the intricacies of hand shapes and gestures. However, the study also
found some limits and possibilities for development. The system's performance on dynamic gestures, particularly for letters like
'J' and 'Z' that entail motion, was substantially worse than for static motions. This underscores the need for more advanced
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIII, Issue XII, December 2024
www.ijltemas.in Page 324
approaches that may incorporate temporal information in gesture identification. Additionally, while the system worked well under
controlled conditions, its robustness in varied real-world environments, such as different lighting conditions and backgrounds,
requires further exploration. The deployment of the system as a web application using Streamlit indicates its potential for
practical, real-world use. However, adapting the system for mobile devices and increasing its real-time processing capabilities
would boost its accessibility and usability for a broader audience. Future research areas should focus on resolving these
constraints, potentially by adding recurrent neural networks or 3D CNNs to better handle dynamic gestures. While obstacles
persist, the existing system represents a promising step towards more inclusive communication technology, with the potential to
substantially enhance the lives of deaf and hard-of-hearing individuals in Nigeria and beyond.
References
1. Alaftekin, M., Pacal, I. & Cicek, K. (2024) Real-time sign language recognition based on YOLO algorithm. Neural
Computing & Applications 36,76097624 https://doi.org/10.1007/s00521-024-09503-6
2. Asonye, E. I., Emma-Asonye, E., & Edward, M. (2018). Deaf in Nigeria: A Preliminary Survey of Isolated Deaf
Communities. Sage Open, 8(2). https://doi.org/10.1177/2158244018786538
3. Asonye, Emmanuel & Emma-Asonye, Ezinne & Edward, Mary. (2020). Linguistic Genocide against Development
of Indigenous Signed Languages in Africa.
4. Cohen, S. (2020). Artificial intelligence and deep learning in pathology. Elsevier Health Sciences.
https://doi.org/10.1016/C2018-0-02465-2
5. Dabwan , Basel & Jadhav, Mukti & Abosaq, Hamad & Olayah, Fekry & Yami, Mohammed & Ali, Yahya.
(2024). Real-time System for Translating American Sign Language to Text Using Robust Techniques.1-
6.10.1109/ICRASET59632.2023.10420110
6. Deshpande, A., Shriwas, A., Deshmukh, V.J., & Kale, S. (2023). Sign Language Recognition System using CNN. 2023
International Conference on Intelligent and Innovative Technologies in Computing, Electrical and Electronics
(IITCEE), 906-911.
7. Eleweke, C. (2002). A Review of Issues in Deaf Education Under Nigeria's 6-3-3-4 Education System. Journal of
deaf studies and deaf education. 7. 74-82. 10.1093/deafed/7.1.74.
8. Gordon, R. G., Grimes, B. F., & Summer Institute of Linguistics. (2005). Ethnologue: Languages of the world (15th ed.).
SIL International.
9. Greenberg, S., Blight, J., & Wong, A. Colour-based Gesture Recognition for American Sign Language via Hidden
Markov Models. University of Waterloo, Canada.
10. Joudaki, S., Mohamad, D., Saba, T., Rehman, A., Al-Rodhaan, M., & Al-Dhelaan, A. (2014). Vision-based sign language
classification: A directional review. IETE Technical Review, 31(5), 383-402.
https://doi.org/10.1080/02564602.2014.961576
11. Karbasi, M., Shah, A., & Landani, Z. (2015). An analysis of vision-based Malaysian sign: A review. International
Journal of Advanced Research in Science, Engineering and Technology, 2(1), 395- 399.
12. Lillo-Martin, D., & Sandler, W. (2006). Sign language and linguistic universals. Cambridge University Press.
Merriam-Webster. (n.d.). Manual alphabet. In Merriam-Webster.com dictionary. Retrieved October 22, 2024, from
https://www.merriam-webster.com/dictionary/manual%20alphabet
13. Padden, C. (2003). How the alphabet came to be used in a sign language. Sign Language Studies, 4(1), 10-33.
https://doi.org/10.1353/sls.2003.0026
14. Pathan, R.K., Biswas, M. and Yasmin, S. (2023). Sign language recognition using the fusion of image and hand
landmarks through multi-headed convolutional neural network. Sci Rep 13, 16975 (2023).
https://doi.org/10.1038/s41598-023-43852-x
15. Paulraj, M. P., Yaacob, S., Azalan, M. S. Z., & Palaniappan, R. (2010). A phoneme based sign language recognition
system using skin color segmentation. 6th International Colloquium on Signal Processing & Its Applications (CSPA), 1-
5. IEEE. https://doi.org/10.1109/CSPA.2010.5545291
16. Oguntimilehin, A., & Balogun, K. (2024). Real-Time Sign Language Fingerspelling Recognition using Convolutional
Neural Network. The International Arab Journal of Information Technology, 21(1).
https://doi.org/10.34028/iajit/21/1/14
17. Shin, Jungpil & Matsuoka, Akitaka & Hasan, Md. Al & Srizon, Azmain. (2021). American Sign Language Alphabet
Recognition by Extracting Feature from Hand Pose Estimation. Sensors (Basel, Switzerland). 21. 10.3390/s21175856.
18. Simon, C. (1982). International hand alphabet charts (2nd ed.). National Association of the Deaf.
19. Swee, T. T., Ariff, A. K., Salleh, S. H., Seng, S. K., & Huat, L. S. (2007). Wireless data gloves Malay sign language
recognition system. 6th International Conference on Information, Communications & Signal Processing, 1-4.
IEEE. https://doi.org/10.1109/ICICS.2007.4449599
20. World Health Organization. (2024). Deafness and hearing loss. https://www.who.int/news-room/fact-
sheets/detail/deafness- and-hearing-loss
21. Zhang, Yanqiong & Jiang, Xianwei. (2024). Recent Advances on Deep Learning for Sign Language Recognition.
Computer Modeling in Engineering & Sciences. 139. 1-10.10.32604/cmes.2023.045731.
