Article Preview
Top1. Introduction
Sign Language is one of the most sophisticated and structured means of communication for those who are hard of hearing and/or mute. Oftentimes, there are situations when they cannot have a conversation with people solely because many are unaware of the different taxonomies involved with Sign Language (SL) gestures. Bidirectional Sign Language Translation (SLT) is the conversion of SL gestures to their corresponding phrases and vice versa. It assists people who are hard of hearing to communicate in the wider world, which predominantly uses spoken language.
Spoken languages are sound-based, whereas sign languages, i.e., gesture-based languages, are concerned with appearance and hence use hand movements, signs in a particular order, and body language to create relevant words. SL communication can have different representations, one being phrase-based and another being character-based. The former representation, which depends on the movement of hands, would need dynamic input in the form of a video for translation, whereas the latter could be translated by taking static images.
SLT systems present today use Machine Learning and Deep Learning methodologies integrated into working application software. These applications are available in the form of data-gloves, mobile and web applications for easy access to the general public.
A detailed overview of the research based on SLT, current developments in the field, and SL barrier concerns are addressed in the study. Different sign languages, modalities, and datasets in SL are addressed and given in a tabular form to make them easier to comprehend. The contributions made by this extensive SLT review paper are presented in the flow chart mentioned in Figure 1.
Figure 1.
Flow chart of contributions
Firstly, the study proposes a review method that will be followed to assess articles; 125 articles were surveyed and the findings were presented in a proposed order in a laconic manner. Further, the drawbacks of specific methods used by some articles were highlighted; neoteric trends of recent papers were accentuated, and future scope and research gaps are identified.
The study analyses 125 articles against research questions to gain significant insights:
- •
Is the SLT system mentioned in the paper one-way or two-way?
- •
Dataset questions that provide more information about the dataset each article uses
- •
What different pre-processing methods are used in the articles?
- •
What different dimensionality reduction techniques are used?
- •
What feature extraction techniques are used in the different SLT systems?
- •
Informative questions about the different models used in each article
- •
What are the existing applications available to the general public and enabling technologies for Sign Language translation and recognition?
- •
What evaluation metrics are used in different articles?
- •
What validation techniques have been used in SLT systems?
These research questions were identified to explore the different methods that are proven to give good accuracy and can be applied in real-time SLT systems. The (World Health Organization, 2021) reports that nearly 2.5 billion people will have some form of hearing loss by 2050, with over a billion young adults who face the risk of permanent hearing loss. Yet, despite extensive study in this sector, SLT's potential applicability for real-time applications has to be fulfilled. Other possible reasons for the gap in theory and practice have been answered (Hao, 2019).