Real-Time Multilingual Sign Language Origin Classification with Temporal Motion Features
Amrul Kayes, Huang Haiyu, Md Redwan Ullah, Chandan Chakma, Mohammad Abdul Oadud · 2025
Sign language varies across regions and cultures, yet automated classification of sign language origin remains challenging due to subtle differences in signing styles. We present TempoSign, a real-time architecture for multilingual sign language origin classification using the SpreadTheSign-Ten (SP10) dataset. Our approach extracts comprehensive features including specialized hand configurations, relative coordinate systems, and temporal motion dynamics through a pipeline that leverages MediaPipe Holistic for landmark extraction. The architecture employs a hybrid CNN-BiLSTM neural network with distinctive feature processing methodology that captures both spatial patterns and temporal relationships between frames. Using robust normalization techniques and dimensionality reduction, we achieve 96.48 % average accuracy across ten sign languages (Bulgarian, German, English, Icelandic, Italian, Lithuanian, Russian, Swedish, Ukrainian, and Chinese) with 2.44× faster inference than state-of-the-art methods (2.26s vs 5.51s). Performance analysis reveals particularly high accuracy for Icelandic (99.0%) and Chinese (98.6%) sign languages, with Russian (90.0 %) having the lowest but still competitive accuracy. While achieving slightly lower accuracy than the best-performing method (2.43% gap), TempoSign's superior computational efficiency makes it highly suitable for real-time applications including live translation systems, assistive technology, and interactive educational platforms.