Idea
Lightweight Bangla Sign Language recognition model and expert-validated dataset enabling real-time smartphone accessibility.
Research Paper
Core Innovation
This paper presents RSBdSL38, a comprehensive expert-validated BdSL image dataset, and a novel lightweight attention-based convolutional network with only 298,470 parameters. The model matches the accuracy of larger pretrained architectures while drastically reducing parameter count and computational cost, enabling efficient on-device inference and practical deployment.
Why It Matters
Bangla Sign Language users face limited access to automated recognition tools due to bulky models and unverified datasets. This solution delivers accurate, efficient recognition on personal devices, expanding educational and service access for deaf and hard-of-hearing communities in Bangladesh. It scales by reducing computational demands, making deployment feasible on commodity smartphones.
Market Size (TAM)
$20–50M TAM for sign language recognition tools in South Asia; $5–10M SAM from educational and accessibility-focused organizations. Driven by increasing smartphone penetration and disability inclusion initiatives.
Potential Customers & Pain Points
- Deaf and hard-of-hearing individuals in Bangladesh – Lack of accessible sign language recognition tools
- Educational institutions for special needs – Need reliable BdSL teaching aids
- Mobile app developers – Require lightweight accurate models for on-device deployment
- NGOs and government agencies – Need scalable communication solutions for disability inclusion.
Business Model
Freemium mobile app offering basic BdSL recognition with premium features for educational institutions and NGOs; licensing dataset and model for integration into third-party accessibility tools.
Competitive Landscape
- Google ASL Recognition
- SignAll
- KinTrans
Implementation Challenges
- Limited awareness and adoption of BdSL technology in target communities
- Challenges in collecting diverse signer data for robust generalization
- Integration with existing educational and communication platforms
Validation Strategy
- Pilot deployment in special-needs schools across Bangladesh
- User feedback collection from deaf and hard-of-hearing individuals
- Benchmarking against public BdSL datasets and signer-independent splits
- Performance and usability testing on commodity smartphones
Research Paper Overview
Toward Deployable Bangla Sign Language Recognition with Expert-Validated Data and a Lightweight Attention-Based Model
Summary
This work introduces RSBdSL38, a large expert-validated dataset of Bangla Sign Language images, and a lightweight attention-based convolutional model optimized for on-device recognition. The model achieves high accuracy comparable to heavyweight pretrained architectures but with significantly fewer parameters and computational costs, enabling real-time deployment on smartphones. The dataset, code, and models are publicly released to support accessible BdSL recognition.