Work

SignBridge AI: Real-Time Sign Language Recognition and Translation System

Year: 2026
Type: preprint
Abstract: SignBridge AI is a camera-based sign-language recognition and translation system designed to lower the communication barrier between Deaf and Hard-of-Hearing (DHH) people and hearing users, to a significant degree. Its application pipeline pulls 75 three-dimensional pose and hand landmarks (225 scalar traits per frame), employs shoulder-centred translation and inter-shoulder distance normalization, and classifies 30-frame sequences using a 1D-CNN and stacked LSTM network, on the whole. The used application supports 44 signs - the 26 ASL letters, numbers 1-5, and 13 conversational or emergency signs; it's offering a Starlette/WebSocket live recognizer and a Streamlit academic dashboard, with async text-to-speech. A three-stage output stabilizer is combining a 0.75 confidence threshold, a 7-of-10 majority vote, and a 15-frame refractory interval. The supplied project benchmark reports 100.0% test accuracy on a held-out set of 60 sequences and a 29.7 ± 3.4 ms end-to-end cycle (33.7 FPS) on an Apple M-series or commodity laptop CPU, the principal processor. One at a time, this paper keeps the original WLASL 10-word research experiment as a separate check - that experiment reported 43.5 ± 8.1% accuracy for its hybrid model and 3.1 frames/s on a two-core Colab CPU; these findings aren't interchangeable, because they refer to diverse datasets, model configurations, and runtime environments. What's different is made unambiguous to avoid exaggerating the proof.
Language: English
Open Access status: green
OpenAlex ID: W7236566480
Cites:
Cited by:
Related to: 10