Zero-Shot Transfer for Sign Language Handshape Recognition

A new framework enables cross-lingual handshape recognition by decomposing signs into five shared phonological features: finger selection, flexion, spread, thumb position, and thumb contact. Using ASL-trained models like SHuBERT, the method achieves 80.0% feature accuracy and 54.5% handshape accuracy when transferred to Catalan Sign Language.