Advanced Deep Learning Architectures for Sign

Language Recognition, Translation, and Multimodal

Abstract—The paper will provide a survey of the recent devel- recognition errors. Other aspects in the real-world setting that
opments in Sign Language Processing (SLP), the field in which increase the discrepancy between the laboratory setting and
there is a paradigm shift to a deep learning architecture based on actual implementation include differences in signing speed,
accessibility equity rather than relying on traditional gloss-based
pipelines. The survey critically examines the major advancements the appearance of the signer, lighting, and clutter on the
in Sign Language Recognition (SLR), Sign Language Translation background. In survey studies conducted by ZainEldin et al.,
(SLT), and Sign Language Production (SLP) and outlines the [2] and Linardakis et al., [6] it is mentioned that when systems
architecture of the evolution from the hybrid CNN-RNN models do not consider non-manual cues (especially facial expressions
to Transformer-based frameworks and Multimodal Large Lan- and lip movement), their performance declines significantly,
guage Models (MLLMs). It is observed that there is a definite
trend of gloss-free translation, which is facilitated by discrete although these cues typically convey critical grammatical
latent tokenization, vector quantization (SignLLM), and segment- information (such as emphasis, negation, and interrogative
aware encoding schemes like Low-SAGE. The survey also looks forms).
at the combination of heterogeneous sensory modalities, such Recent developments In Sign Language Processing (SLP),
as RGB and depth streams, skeletal grids, neuromorphic event
this has been the shift towards deep learning-based methods
cameras (VECSL), and electromyography (sEMG), to overcome
issues such as motion blur and occlusion. Specific attention is away from hand-designed features and Hidden Markov Mod-
given to linguistic inclusivity by examining large-scale standards els [2], [3]. Although initial networks based on CNNs and
of under-resourced sign languages, such as Indian (FDMSE- LSTMs enhanced the spatiotemporal modeling [7], [8], the
ISL), Persian, Bangla (BdSLW401), and Russian (Logos). This recent studies show a growing adoption of Transformer-based
survey summarizes the ongoing challenges of the deployment of
networks and Large Language Models (LLMs) [9], [10]. Such
continuous recognition, retrieval-enhanced translation (OVSLT),
and diffusion-based avatar synthesis (TDM, SOKE) and forms frameworks as SignLLM and Sign2GPT use the linguistic
the outlook of how powerful, bi-directional, and human-friendly knowledge of frozen LLMs to facilitate gloss-free translation
sign language foundation models can be built. [11], [12], whereas Retrieval-Augmented Generation (RAG)
Index Terms—Sign Language Processing, Gloss-Free Transla- and context-aware models use background and sentence-level
tion, Deep Learning Architectures, Multimodal Learning, Acces-
context to alleviate ambiguity [11], [12]. Meanwhile, the
sibility Equity
research focus on accessibility has become one of the topic
areas, as the dominance of high-resource sign languages is now
I. I NTRODUCTION
discussed [1]. Projects like FDMSE-ISL (Patra et al.) [13], the
Sign languages are fully natural languages that have their Persian Sign Language dataset provided by Rastgoo et al., [14]
own grammar and vocabulary, the main users of which are the or research on Bangla, Marathi, and Russian Sign Languages
Deaf and Hard-of-Hearing (DHH) community. In contrast to are trying to minimize this inequality [15]–[17]. Moreover,
spoken languages, which have linear acoustic patterns, sign the bi-directional communication systems are being worked
languages are spatial-visual and depend upon simultaneous on as Wilkins et al [18]. point out that real-world usefulness
application of both the manual aspects of language (i.e., hand is not only based on accuracy but also efficiency and usability
shape, orientation, movement) and non-manual representation compared with a human interpreter.
(i.e., facial expression, head movement, body posture, etc.)
[1]. Such a multimodal character adds greatly to the complex- II. LITERATURE SURVEY
ity of automatic sign language processing in comparison to
spoken/written language systems [2]–[4]. To provide a structured overview of representative deep
One of the most difficult issues involves continuous signing, learning architectures across recognition, translation, and pro-
where the signs lack boundaries. According to Geetha et al., duction tasks, Table I summarizes key models, their target
[5] and Toshpulatov et al., [3] such co-articulation causes many tasks, and core methodological contributions. This taxonomy
highlights the architectural diversity and evolving trends in be a major component in the learning of unsegmented visual
modern Sign Language Processing research. inputs onto label sequences [3]. Wang et al. came up with
Multi-Scale Spatial-Temporal Feature Enhancement Network
TABLE I (STNet) but added a spatial resonance module based on the op-
S UMMARY OF R EPRESENTATIVE D EEP L EARNING A RCHITECTURES FOR timal transport theory to extract shared representations among
S IGN L ANGUAGE P ROCESSING
adjacent frames [25]. This gives definite performance increases
Author Method Dataset Strengths
on PHOENIX14 benchmark. Ahn et al. scaled SlowFast archi-
tectures to CSLR, (with) two processing pathways which run
Ahn et al. (2021) SlowFastSign CSLR Corpora Captures fast and slow motion dynamics
at different temporal resolutions to simulate both the high-level
Wang et al. (2022) STNet PHOENIX14 Multi-scale spatio-temporal modeling
spatial semantic and fine-grained motion behavior with respect
Miah et al. (2022) GmTC Multi-language SLR Robust signer-invariant recognition to each other [26]. Ranjbar and Taheri followed this with the
Zuo et al. (2022) SOKE SLP Corpora Structured autoregressive sign generation intra and inter gloss attention mechanism to make explicit
Tang et al. (2022) Discrete2Cont SLP Datasets Smooth transition between sign units
the inter frame but also inter gloss frame dependencies [27]–
[29]. This builds up the knowledge of sequences. Geetha et al.
Gong et al. (2023) SignLLM SLT Benchmarks Gloss-free translation using LLMs
however are practical, building the SignFlow framework that
Wong et al. (2023) Sign2GPT SLT Datasets Pseudo-gloss pretraining improves translation
pairs the dynamic video down-sampling with a transformer
Park et al. (2023) Low-SAGE SLT Datasets Efficient segment-aware tokenization architecture that does not use positional encoding to efficiently
Kim et al. (2023) MMSLT Multimodal SLT Integrates motion and text modalities execute that model within real-time constraints [5], [30].
He et al. (2023) TDM Pose-based SLP Anatomically consistent diffusion generation
C. Gloss-Free Translation Architectures
The emphasis of gloss-free translation as the subject of the
A. Isolated Sign Recognition Systems new research has a paradigmatic aspect, as it replaces the time-
Isolated Sign Language Recognition (ISLR) aims at rec- consuming process of dissecting the translation into glosses
ognizing single, segmented signs and forms the basis of by hand with the acquisition of direct matching of the video
continuous recognition systems [2], [3]. Recent studies have to text [1], [9]. Similarly, Gong et al. propose the SignLLM
mostly focused on lightweight and efficient architectures to (Translation) model that constructs a language-like model of
provide the ability to deploy them on resource-constrained the sign videos by vector-quantized modules and codebook
devices like smartphones [19]. A mobile-friendly framework alignment to engage with a frozen Large Language Model
suggested by Ghanimi et al. is the combination of MediaPipe- efficiently [9], [31]. On this line, Wong et al. present the
based skeletal tracking and particle filtering with a hybrid Sign2GPT technique, whereby the pseudo-glosses generated
CNN -BiLSTM model, providing a high recognition accuracy by natural language processing applications are used to pre-
with low computation requirements [19]. Likewise, Kothadiya train sign encoders to ensure that translation is a more viable
et al. introduced Hybrid InceptionNet that is more efficient in process without gloss annotation [12]. Kim et al. continued
training due to the optimized network connectivity, and the this front with Multimodal Large Language Models in their
reported accuracy is 98.46 percent on single sign recognition MMSLT framework that produced a textual description of sign
[20]. movements and used this with image features to fill in the
In addition to the traditional CNN-based techniques, graph- intermodal translation gap between videos and language inputs
based methods have become dominant models used to learn [10], [32]. Furthermore, Park et al. introduced a segment-
the hand and body movements [6], [21]. Patra et al., used aware tokenization scheme called Low-SAGE that introduced
Hierarchical Windowed Graph Attention Networks (HWGAT) an effective scaling-up strategy that had a reduction in the
to better capture skeletal motion, which showed good results segment length of up to 50 percent in transformer-based
on a large-scale dataset (FDMSE-ISL) of skeletal motion [13]. models [33]. Still in the gloss-free translation (the same),
Simultaneously, Baihan et al. investigated the application of Tan et al. came up with a combined translation model to
meta-heuristic optimization by incorporating the Hippopota- see translation across the translation of ten different sign
mus Optimization algorithm in conjunction with the Pathfinder languages that combined token-level language identification
Algorithm in the performance improvement of a CNN-LSTM- to solve inter-language conflicts [34], [35].
based ISLR model, reflecting the current increase of interest
D. Sign Language Production Systems
in the optimization-related performance enhancements of the
research on ISLR [22], [23]. Sign Language Production (SLP) is regarded as the inverse
of recognition, and it is aimed at making videos or pose
B. Continuous Sign Language Recognition sequences considering text input [3]. New developments in
The objectives of CSLR are the identification of a sequence diffusion models in recent years have had a tremendous effect
of signs in unsegmented video streams that consist of con- in this field in producing high-quality pose sequences with
tinuous streams of video, and this is basically a difficult task a temporal smoothing effect [36], [37]. In order to solve
because there are no explicit edges between adjacent signs this significant problem of discrete-to-continuous transition,
[2], [24]. In this aspect the CTC loss mechanism continues to Tang et al. exploited a conditional diffusion model to generate
continuous transition frames during the process of changing home automation system by using Leap Motion Controllers
one sign into another [36]. to achieve coherence between them [43]. Of sensor-based
In order to obtain improved performance in pose sequence systems, Soosaimariyan et al. demonstrated effective use of
generating, He et al. added a Text-Driven Diffusion Model the sensor-centric Surface Electromyography technique with
named TDM, where the text-based conditions were employed Generalized Deep Infomax methods to classify pure hand
in eliminating noise during pose sequence generation with an signs by using the sEMG signals alone, providing a nonvisual
additional constraint of orientation of the bones to achieve methodology of understanding signs [44]. This was enhanced
more realistic anatomical structure [37]. As an alternative by Habib et al., who added an ensemble multi-stream CNN
avatar-based perspective, Fang et al. constructed a multilingual model to strongly combine the two skeletal and data-level
production system, named Sign LL-Production, but with the modalities to aid real-time identification [45]. The original
ability to alternate between eight different sign languages sensor-driven recognition systems like the surveys conducted
optimized by reinforcement learning goals to achieve a high- by authors Shin et al. and other surveys conducted by authors
quality production [11]. Hashi et al., all identify the hybrid mode methods which
As a supplementary but also interesting paper, Zuo et is a combination of visual channels and sensor-driven input
al. proposed SOKE, a work that explains the generation of channels is always superior to the purely visual input channels
signs through an autoregressive model considering discrete in the more complex environments [46]–[48].
tokens with a decoupled tokenizer to various sections of a
human being and a dictionary-based search of high accuracy III. DATASETS AND EVALUATION METRICS
in generating works [38]. In other work published by Ngo et Datasets and evaluation metrics play a critical role in
al., a system is developed that converts the Vietnamese texts enabling fair, reproducible, and comparative analysis of sign
into 3D pose motion of digital avatars, and another aspect language processing systems. Table II provides a consolidated
where it is used is in educational education through a sign overview of representative benchmark and regional datasets,
language production system [39]. highlighting their linguistic coverage, recognition type, and
scale, with particular emphasis on under-resourced sign lan-
E. Regional and Multilingual Systems guages.
The search for the accessibility equity of the sign lan-
guage has resulted in the remarkable enhancement of the TABLE II
establishment of the regional sign language systems that pay OVERVIEW OF R EPRESENTATIVE S IGN L ANGUAGE DATASETS
more attention to the accommodation of the superior linguistic
typologies [1], [24]. The GmTC model was suggested by Dataset Language Type Scale
Miah et al. as a graph-based structure of multicultural sign
FDMSE-ISL Indian (ISL) Isolated 40k videos / 2,002 signs
language recognition that has reached a respectable level of
performance in five languages, namely, Korean, American, and Logos Russian (RSL) Isolated 200k videos / 2,863 signs
Bengali [40]. In the case of Indian Sign Language, Sharma et
al. and Geetha et al. incorporated the continuous recognition BdSLW401 Bangla (BdSL) Isolated 401 classes
frameworks, which are sensitive to the syntactic nature of the
Indian Sign Language [5]. Aadeeb et al., in the case of Bangla Persian SL Persian (PSL) Continuous 10k videos / 100 sentences

Sign Language, suggested BdSLW401 and employed Relative VECSL Chinese (CSL) Continuous RGB + Event streams
Quantized Embedding to manage the variability of signers
successfully [15]. Rastgoo et al. further published a large-scale SP-10 Multilingual Isolated 10 languages
continuous Persian Sign Language dataset, which enabled
sentence-level research to take place in the linguistic setting UMANG-ISL Indian (ISL) Continuous Sentence-level corpus
traditionally under-resourced [14]. Modi et al. also handled
the Marathi Sign Language through edge detection with CNN-
based models to enhance the accessibility of regional users in A. Benchmark Datasets
India [16]. There should be standard benchmark datasets to enable
fair and reproducible comparisons between a number of
F. Multimodal Gesture Foundations approaches to sign language processing [1]. The RWTH-
In many cases, effective SLP is not limited to mere assess- PHOENIX-Weather 2014T database is still regarded as one
ment of the RGB video inputs since the fusion of multimodal of the gold standards of continuous sign language translation,
can greatly enhance the stability of different environments which includes both data in the German Sign Language
[2], [41]. Wang et al. suggested a VECSL dataset, which with annotated gloss and text in spoken language [3], [24].
also includes RGB inputs alongside event camera signals to Isolated sign recognition datasets like WLASL—World Level
overcome high-speed blurs on signs as a result of motion American Sign Language—and MS-ASL are highly utilized
during capture [42]. Bhiri et al. introduced the 2MLMD in the context of isolated sign recognition due to their large
dataset, which deploys skeletal and depth modalities on a vocabulary and training under natural conditions, and they
also present realistic problems such as occlusion and chang- standards between datasets, including the difference between
ing light conditions [2], [6]. More recently, American Sign OpenPose and MediaPipe keypoint representations, also limit
Language translation received considerable attention in regard cross-dataset transfer learning, as noted by Linardakis et al
to the How2Sign dataset, which was a large-scale, multimodal [6], [46].
resource with a specific focus on facilitating research on open-
domain translation [9], [49]. B. Multilingual and Regional Coverage
Despite the current developments in the field of accessibility
B. Regional Datasets equity, there are still major differences. The majority of the
In the more recent development, in their works of research, current foundational models still deal with high-resource sign
with a target on the linguistic area, they are trying to draw languages, like ASL and German Sign Language (DGS) [9],
datasets to handle an improvement of the equity of acces- [24]. Marathi and Bangla sign language research papers point
sibility [1]. As an example, the FDMSE-ISL dataset includes at how low-resource areas are disadvantaged such that even
more than 40000 videos (2002 signs in Indian Sign Language), simple isolated sign collections are few [15], [16]. Tan et
and they are represented with atomic-level and composite-level al. also raise the problem of language conflict in unified
annotations [13]. Ovodov et al. introduced a Logos dataset of multilingual models, in which the grammatical composition of
RSL, which, despite being massively crowdsourced, generated one sign language has a detrimental impact on the learning of
massive volumes of videos comprising nearly 200,000 videos another [34]. Consequently, generally universal sign language
and was specifically constructed to answer visually similar foundation models are restricted, in that they do not scale to
signs [17]. The purpose behind the use of the BdSLW401 new sign languages unless heavily fine-tuned [10], [38], [51].
dataset is to use word-level Bangla Sign Language recognition
in a low-resource environment [15]. On the same note at the C. Real-Time Deployment Barriers
Middle Eastern level, detailed contributions made on Persian The use of high-performance deep learning models on
Sign Language by Rastgoo et al. have leveled a valuable source edge devices to serve Deaf and Hard-of-Hearing users is
of C-SLR research in the region [14]. still not well-developed because of the high computational
complexity of such architectures as Transformers and Graph-
C. Evaluation Metrics based models [6], [24]. According to the research, by Ghanimi
Sign language system assessment is based on computer et al. and Geetha et al., it is possible to use lightweight
vision and natural language processing measures [3], [24]. models on smartphones, but there is a necessity to carefully
In Continuous Sign Language Recognition (CSLR), Word balance recognition accuracy, latency, and user experience to
Error Rate (WER) is still the most important, because it is implement it in practice [5], [19]. Wilkins et al. state that we
used to measure the edit distance between a predicted and need low latency to have usable bi-directional communication
reference gloss sequence [2]. BLEU and ROUGE are typically and this is something that is overlooked in accuracy-driven
employed in the evaluation of the quality of generated text studies [18]. Although the Resampling Hook model allows
in Sign Language Translation (SLT) [9]. Nevertheless, these local modifications without recalculating the entire model
measures tend not to reflect the richer semantic overtones of computation, efficiency-based methods have not been widely
sign language and can be used with interest in the strategy of used. [33], [34], [52], [53].
evaluation using LLM, as Jang et al. put it through [50]. In
the case of Sign Language Production (SLP), Frechet Inception D. Practical Accessibility Gaps
Distance (FID) and DTW-MJE are the most common metrics It is not all about technological sophistication that makes it
to assess the visual realism and motion accuracy of reference practical to use [1]. The thorough usability study conducted by
sequences to pose quality and temporal fidelity [36], [37]. Wilkins et al. proved that even the most accurate systems can
be refused, unless they can build user trust or transparency in
IV. RESEARCH GAPS AND CHALLENGES their operation [18]. In addition, the various available methods
A. Data Scarcity and Annotation Issues fail in practice, in the field, and recent robustness studies have
Although the list of the existing datasets is constantly shown declining performance in low-light and under occulsion
increasing, large amounts of high-quality annotated data, espe- [54], [55]. Another weakness is the inadequate representation
cially with non-Western languages, are still necessary in con- of non-manual elements, including facial expressions, resulting
tinuous sign language processing [1], [24]. Gloss annotation, in stiff or rather robot-like avatars incapable of expressing the
as Toshpulatov et al. and ZainEldin et al. cite, is very time- emotional and grammatical subtext to effective communication
intensive and requires specific linguistic knowledge, which in natural sign language communication [11], [39], [56].
is a significant bottleneck [2], [3]. Nevertheless, in spite of V. FUTURE DIRECTIONS
the fact that gloss-free methods like SignLLM are developed
to be less dependent on this dependency, they also use large A. Foundation Models and Transfer Learning
volumes of data that are paired with video-text to provide good To progress sign language processing, frozen Large Lan-
generalization [9], [11]. Moreover, cross-dataset transfer learn- guage Models (LLMs) should be adapted, and multimodal
ing is also constrained because inconsistencies in annotation LLMs that do not require gloss-free translation should be
trained [9], [10]. SignLLM and Sign2GPT approaches indicate and diffusion generation are displaying significant potential as
that the main difficulty is to match the visual sign repre- far as future systems are concerned. The technical success of
sentations with textual semantics [11], [12]. Future studies such technologies will eventually hinge on how much they
ought to focus on self-supervised learning to create large- are adopted by these communities, such as the Deaf and hard-
scale Sign Language Foundation Models similar to GPT-4 in of-hearing people across the world, depending on factors of
natural language processing [34], [38]. This type of model usability and trustworthiness.
would allow few-shot transfer to low-resource sign languages,
which is a significant leap towards providing equal access to R EFERENCES
digital resources [1], [24].
[1] Fox, N., Woll, B., & Cormier, K. (2025). Best practices for sign language
B. Bidirectional Systems and Avatars technology research. Universal Access in the Information Society, 24(1),
69-77.
The accessibility will rely on the ability to provide two-
[2] ZainEldin, H., Gamel, S. A., Talaat, F. M., Aljohani, M., Baghdadi,
way communication between Deaf and hearing people in a N. A., Malki, A., ... & Elhosseini, M. A. (2024). Silent no more:
fast and effective way [1], [18]. Therefore, the new systems a comprehensive review of artificial intelligence, deep learning, and
should combine powerful Sign Language Recognition (SLR) machine learning in facilitating deaf and mute communication. Artificial
Intelligence Review, 57(7), 188.
to support the Deaf users with good Sign Language Production [3] Toshpulatov, M., Lee, W., Jun, J., & Lee, S. (2025). Deep learning
(SLP) to support the hearing users [3], [57]. Studies are pathways for automatic sign language processing. Pattern Recognition,
supposed to consider more human-oriented designs including 111475.
[4] Al Abdullah, B. A., Amoudi, G. A., & Alghamdi, H. S. (2024).
the framework of user-editable avatar suggested by which lets Advancements in sign language recognition: A comprehensive review
users inspect and enhance generated signs, thus, enhancing the and future prospects. IEEE Access, 12, 128871-128895.
level of trust and ease [58]. The further development of the [5] Geetha, M., Aloysius, N., Somasundaran, D. A., Raghunath, A., &
Nedungadi, P. (2025). Towards Real-Time Recognition of Continuous
diffusion-based generation models is needed to achieve photo- Indian Sign Language: A Multi-Modal Approach using RGB and Pose.
realistic sign videos that maintain time flow and display non- IEEE Access.
manual markers correctly [36], [37], [57]–[59]. [6] Linardakis, M., Varlamis, I., & Papadopoulos, G. T. (2025). Sur-
vey on hand gesture recognition from visual input. arXiv preprint
C. Community-Driven Validation arXiv:2501.11992.
[7] Huang, J., & Chouvatut, V. (2024). Video-based sign language recogni-
The conventional technical measurements, Word Error Rate tion via resnet and lstm network. Journal of Imaging, 10(6), 149.
(WER) and BLEU, do not entirely reflect the usefulness of [8] Paul, S. K., Walid, M. A. A., Paul, R. R., Uddin, M. J., Rana, M. S.,
Devnath, M. K., ... & Haque, M. M. (2024). An Adam based CNN
sign language systems in a real-world scenario [24], [50]. The and LSTM approach for sign language recognition in real time for deaf
future studies should thus focus on human-based appraisal sys- people. Bulletin of Electrical Engineering and Informatics, 13(1), 499-
tems, especially those that engage Deaf community members 509.
[9] Gong, J., Foo, L. G., He, Y., Rahmani, H., & Liu, J. (2024). Llms
as suggested by Wilkins et al., and the best practice outlined are good sign language translators. In Proceedings of the IEEE/CVF
in the scale of the System Usability Scale (SUS) and user Conference on Computer Vision and Pattern Recognition (pp. 18362-
confidence by Fox et al., [1], [18]. Simultaneously, the trend 18372).
[10] Kim, J., Jeon, H., Bae, J., & Kim, H. Y. (2025). Leveraging the power
of dataset development is becoming more decentralized and of mllms for gloss-free sign language translation. In Proceedings of the
community-based, with efforts like Logos and FDMSE-ISL IEEE/CVF International Conference on Computer Vision (pp. 21048-
being an example of ethical data collection methods that focus 21058).
[11] Fang, S., Chen, C., Wang, L., Zheng, C., Sui, C., & Tian, Y. (2025).
on the issue of informed consent and data privacy [2], [13], Signllm: Sign language production large language models. In Proceed-
[17], [60]. ings of the IEEE/CVF International Conference on Computer Vision
(pp. 6622-6634).
VI. CONCLUSION [12] Wong, R., Camgoz, N. C., & Bowden, R. (2024). Sign2GPT: Leveraging
large language models for gloss-free sign language translation. arXiv
The Survey of the Development of Sign Language Process- preprint arXiv:2405.04164.
ing study offered information on the accelerated advancement [13] Patra, S., Maitra, A., Tiwari, M., Kumaran, K., Prabhu, S., Punyesh-
of Sign Language Processing (SLP) technology at an initial warananda, S., & Samanta, S. (2024). Hierarchical windowed graph
attention network and a large scale dataset for isolated indian sign
level with SLP research relying on glossing, convoluted neural language recognition. arXiv preprint arXiv:2407.14224.
networks (CNNs), and related technologies and then more [14] Rastgoo, R. (2025). A Persian Continuous Sign Language Dataset.
advanced systems relying on large language models (LLMs), Journal of AI and Data Mining, 13(1), 95-105.
[15] Rubaiyeat, H. A., Youssouf, N., Hasan, M. K., & Mahmud, H.
which accept a broader range of input types, including camera (2025). BDSLW401: Transformer-based word-level Bangla sign lan-
images, EMG sensor data, and event-based vision (activated guage recognition using relative quantization encoding (RQE). arXiv
on capture). The variety of input types employed and the preprint arXiv:2503.02360.
[16] Mali, Y. K. (2025). Marathi sign language recognition methodology
ongoing expansion of regional and low-resource data sets are using Canny’s edge detection. Sādhanā, 50(4), 268.
pointers to a more organized approach in order to promote [17] Ovodov, I., Surovtsev, P., Kvanchiani, K., Kapitanov, A., & Nagaev,
access equity on a global scale. Although both limitations, A. (2025). Logos as a Well-Tempered Pre-train for Sign Language
both the capability to implement SLP in real-time scenarios Recognition. arXiv preprint arXiv:2505.10481.
[18] WILKINS, N., YU, Z., KELLY, N., MURPHY, D., & PAYANO, Y.
and the capability to fully fidelize the linguistic richness of Comparing Sign Language AI and Human Interpreters: Evaluating
sign languages, remain, recent advances in foundation models Efficiency for Greater Social Sustainability.
[19] Ghanimi, H. M., Sengan, S., Sadu, V. B., Kaur, P., Kaushik, M., and general deep learning network. IEEE Open Journal of the Computer
Alroobaea, R., ... & Dadheech, P. (2024). An open-source MP+ CNN+ Society, 5, 144-155.
BiLSTM model-based hybrid model for recognizing sign language on [41] Kyranou, I., Szymaniak, K., & Nazarpour, K. (2025). EMG dataset for
smartphones. International Journal of System Assurance Engineering gesture recognition with arm translation. Scientific Data, 12(1), 100.
and Management, 15(8), 3794-3806. [42] Wang, X., Li, Y., Wang, F., Jiang, B., Wang, Y., Tian, Y., ... & Luo,
[20] Kothadiya, D. R., Bhatt, C. M., Kharwa, H., & Albu, F. (2024). Hybrid B. (2025). Sign language translation using frame and event stream:
InceptionNet based enhanced architecture for isolated sign language Benchmark dataset and algorithms. arXiv preprint arXiv:2503.06484.
recognition. IEEE Access, 12, 90889-90899. [43] Bhiri, N. M., Ameur, S., Jegham, I., Alouani, I., & Ben Khalifa, A.
[21] Miah, A. S. M., Hasan, M. A. M., Nishimura, S., & Shin, J. (2024). (2025). 2MLMD: Multi-modal leap motion dataset for home automation
Sign language recognition using graph and general deep neural network hand gesture recognition systems. Arabian Journal for Science and
based on large scale dataset. IEEE Access, 12, 34553-34569. Engineering, 50(10), 7511-7535.
[22] Baihan, A., Alutaibi, A. I., Alshehri, M., & Sharma, S. K. (2024). Sign [44] Soosaimariyan, M. V., Allasi, H. L., Chitra, K., & Gnanadurai, J. B.
language recognition using modified deep learning network and hybrid (2025). Enhanced EMG-Based Hand Gesture Recognition by Using
optimization: a hybrid optimizer (HO) based optimized CNNSa-LSTM Generalized Deep Infomax Networks. Journal of Sensors, 2025(1),
approach. Scientific Reports, 14(1), 26111. 9496890.
[23] Hax, D. R. T., Penava, P., Krodel, S., Razova, L., & Buettner, R. (2024). [45] Habib, M. K., Yusuf, O., & Moustafa, M. (2025). Skeleton-Based Real-
A novel hybrid deep learning architecture for dynamic hand gesture Time Hand Gesture Recognition Using Data Fusion and Ensemble
recognition. IEEE Access, 12, 28761-28774. Multi-Stream CNN Architecture. Technologies, 13(11), 484.
[24] Khan, A., Jin, S., Lee, G. H., Arzu, G. E., Nguyen, T. N., Dang, L. M., [46] Shin, J., Miah, A. S. M., Kabir, M. H., Rahim, M. A., & Al Shiam,
... & Moon, H. (2025). Deep learning approaches for continuous sign A. (2024). A methodological and structural review of hand gesture
language recognition: A comprehensive review. IEEE Access. recognition across diverse data modalities. IEEE Access.
[25] Wang, Z., Li, D., Jiang, R., & Okumura, M. (2025). Continuous sign [47] Hashi, A. O., Hashim, S. Z. M., & Asamah, A. B. (2024). A systematic
language recognition with multi-scale spatial-temporal feature enhance- review of hand gesture recognition: An update from 2018 to 2024. IEEE
ment. IEEE Access. Access, 12, 143599-143626.
[26] Ahn, J., Jang, Y., & Chung, J. S. (2024, April). Slowfast network [48] Al Mudawi, N., Ansar, H., Alazeb, A., Aljuaid, H., AlQahtani, Y.,
Algarni, A., ... & Liu, H. (2024). Innovative healthcare solutions: robust
for continuous sign language recognition. In ICASSP 2024-2024 IEEE
International Conference on Acoustics, Speech and Signal Processing hand gesture recognition of daily life routines using 1D CNN. Frontiers
(ICASSP) (pp. 3920-3924). IEEE. in Bioengineering and Biotechnology, 12, 1401803.
[49] Mohammadi, Z., Akhavanpour, A., Rastgoo, R., & Sabokrou, M. (2024).
[27] Ranjbar, H., & Taheri, A. (2025). Continuous sign language recognition
Diverse hand gesture recognition dataset. Multimedia Tools and Appli-
using intra-inter gloss attention. Multimedia Tools and Applications, 1-
cations, 83(17), 50245-50267.
19.
[50] Jang, Y., Raajesh, H., Momeni, L., Varol, G., & Zisserman, A. (2025).
[28] Guan, M., Wang, Y., Ma, G., Liu, J., & Sun, M. (2025). MSKA: Multi- Lost in translation, found in context: Sign language translation with
stream keypoint attention network for sign language recognition and contextual cues. In Proceedings of the Computer Vision and Pattern
translation. Pattern Recognition, 165, 111602. Recognition Conference (pp. 8742-8752).
[29] Zhu, Q., Li, J., Yuan, F., & Gan, Q. (2024). Multiscale temporal network [51] Zuo, R., & Mak, B. (2024). Improving continuous sign language recog-
for continuous sign language recognition. Journal of Electronic Imaging, nition with consistency constraints and signer removal. ACM Trans-
33(2), 023059-023059. actions on Multimedia Computing, Communications and Applications,
[30] Srivastava, S., Singh, S., Pooja, & Prakash, S. (2024). Continuous 20(6), 1-25.
sign language recognition system using deep learning with MediaPipe [52] Hwang, E. J., Cho, S., Lee, J., & Park, J. C. (2025, April). An Efficient
holistic. Wireless Personal Communications, 137(3), 1455-1468. Gloss-Free Sign Language Translation Using Spatial Configurations and
[31] Wang, A., Li, J., Luan, W., & Pan, L. (2025). OVSLT: Advancing Sign Motion Dynamics with LLMs. In Proceedings of the 2025 Conference
Language Translation with Open Vocabulary. Electronics, 14(5), 1044. of the Nations of the Americas Chapter of the Association for Compu-
[32] Zhao, R., Zhang, L., Fu, B., Hu, C., Su, J., & Chen, Y. (2024, March). tational Linguistics: Human Language Technologies (Volume 1: Long
Conditional variational autoencoder for sign language translation with Papers) (pp. 3901-3920).
cross-modal alignment. In Proceedings of the aaai conference on artifi- [53] Cui, C., Sunar, M. S., & Su, G. E. (2025). Deep vision-based real-time
cial intelligence (Vol. 38, No. 17, pp. 19643-19651). hand gesture recognition: a review. PeerJ Computer Science, 11, e2921.
[33] Low, J. H., Sincan, O. M., & Bowden, R. (2025). SAGE: Segment-Aware [54] Alaftekin, M., Pacal, I., & Cicek, K. (2024). Real-time sign language
Gloss-Free Encoding for Token-Efficient Sign Language Translation. In recognition based on YOLO algorithm. Neural Computing and Appli-
Proceedings of the IEEE/CVF International Conference on Computer cations, 36(14), 7609-7624.
Vision (pp. 4952-4961). [55] Wu, X., Li, H., Luo, Y., Cheng, X., Zhuang, X., Cao, M., & Fu, K.
[34] Tan, S., Miyazaki, T., & Nakadai, K. (2025). Multilingual Gloss- (2024, September). Uncertainty-aware sign language video retrieval with
free Sign Language Translation: Towards Building a Sign Language probability distribution modeling. In European Conference on Computer
Foundation Model. arXiv preprint arXiv:2505.24355. Vision (pp. 390-408). Cham: Springer Nature Switzerland.
[35] Hu, L., Feng, W., Gao, L., Liu, Z., & Wan, L. (2024). Corrnet+: Sign [56] Wolfe, R., Braffort, A., Efthimiou, E., Fotinea, E., Hanke, T., &
language recognition and translation via spatial-temporal correlation. Shterionov, D. (2025). Special issue on sign language translation and
arXiv preprint arXiv:2404.11111. avatar technology. Universal Access in the Information Society, 24(1),
[36] Tang, S., He, J., Cheng, L., Wu, J., Guo, D., & Hong, R. (2025). 1-3.
Discrete to Continuous: Generating Smooth Transition Poses from Sign [57] Najib, F. M. (2025). Sign language interpretation using machine learning
Language Observations. In Proceedings of the Computer Vision and and artificial intelligence. Neural Computing and Applications, 37(2),
Pattern Recognition Conference (pp. 3481-3491). 841-857.
[37] He, J., Wang, X., Zhang, R., Tang, S., Wang, Y., & Cheng, L. (2025). [58] Li, Y. (2025). Design an Editable Speech-to-Sign-Language Trans-
Text-driven diffusion model for sign language production. arXiv preprint former System: A Human-Centered AI Approach. arXiv preprint
arXiv:2503.15914. arXiv:2506.14677.
[38] Zuo, R., Potamias, R. A., Ververas, E., Deng, J., & Zafeiriou, S. (2025). [59] Tasyurek, S. M., Kiziltepe, T., & Keles, H. Y. (2025). Disen-
Signs as Tokens: A retrieval-enhanced multilingual sign language gen- tangle and regularize: Sign language production with articulator-
erator. In Proceedings of the IEEE/CVF International Conference on based disentanglement and channel-aware regularization. arXiv preprint
Computer Vision (pp. 23806-23816). arXiv:2504.06610.
[39] Ngo, T. D., Nguyen, D. H. L., & Luong, H. L. (2025). Sign Language [60] Kavediya, H., Nayak, V., Sharma, B., & Palaniappan, B. (2025).
Representation using Virtual Characters with 3D Animation. VNU IsoSignVid2Aud: Sign Language Video to Audio Conversion without
Journal of Science: Computer Science and Communication Engineering, Text Intermediaries. arXiv preprint arXiv:2510.07837.
41(1 (2025)).
[40] Miah, A. S. M., Hasan, M. A. M., Tomioka, Y., & Shin, J. (2024).
Hand gesture recognition for multi-culture sign language using graph