GestureTalk: Real-Time Sign Language Recognition Using Deep Learning

GestureTalk+, a real-time Indian Sign Language (ISL) recognition and speech synthesis system designed for commodity Android devices, combining MediaPipe landmark extraction with lightweight temporal models (BiLSTM and Transformer encoders) and quantized on-device inference to achieve sub-300 ms end-to-end latency under subject-independent evaluation.[1]The system targets accessibility in education and healthcare, addressing robustness under variable lighting, backgrounds, and camera viewpoints, and includes bilingual text-to-speech to support Hindi/English output. Experiments on a 26-letter static ISL set and 20 common dynamic words demonstrate macro-F1 of 0.93 and 0.89 respectively, outperforming SVM and CNN-TCN baselines while maintaining 30–45 FPS with NNAPI delegates on mid-range hardware. Ablation studies show that temporal attention and landmark normalization improve confusion cases (e.g., M vs N), and int8 quantization preserves accuracy while reducing compute and power. The paper contributes a reproducible edge-AI pipeline, signer-independent protocols, and deployment telemetry for latency and battery use

Paper

Full text

PDF

GestureTalk: Real-Time Sign Language Recognition Using Deep Learning

Semantic Scholar · 2026

Abstract

GestureTalk+, a real-time Indian Sign Language (ISL) recognition and speech synthesis system designed for commodity Android devices, combining MediaPipe landmark extraction with lightweight temporal models (BiLSTM and Transformer encoders) and quantized on-device inference to achieve sub-300 ms end-to-end latency under subject-independent evaluation.[1]The system targets accessibility in education and healthcare, addressing robustness under variable lighting, backgrounds, and camera viewpoints, and includes bilingual text-to-speech to support Hindi/English output. Experiments on a 26-letter static ISL set and 20 common dynamic words demonstrate macro-F1 of 0.93 and 0.89 respectively, outperforming SVM and CNN-TCN baselines while maintaining 30–45 FPS with NNAPI delegates on mid-range hardware. Ablation studies show that temporal attention and landmark normalization improve confusion cases (e.g., M vs N), and int8 quantization preserves accuracy while reducing compute and power. The paper contributes a reproducible edge-AI pipeline, signer-independent protocols, and deployment telemetry for latency and battery use

Similar papers

© 2026 NYSGPT2525 LLC