Google DeepMind has built SL2T, a sign-language-to-text model now shipping inside Google products for Deaf and hard-of-hearing users. The model translates continuous signed language directly into written text, moving beyond the isolated gesture recognition that has stalled this field for years.
The technical architecture is what warrants a full read. DeepMind did not bolt a translation layer onto an existing speech model. SL2T was built to handle the spatial and temporal grammar of sign language natively, which means it processes handshape, movement, and non-manual markers like facial expression as simultaneous inputs, not sequential ones.
The article details the dataset construction and the specific deployment choices Google made to get from research model to a feature real users can access today. Those decisions, not the benchmark numbers, are where the actual lessons are.
[READ ORIGINAL →]