Sign Language Translation
Overview
Sign language translation turns continuous video of signing into text, requiring models to capture hand motion, facial expression, timing, and sign-language grammar. Systems may predict intermediate glosses or translate directly from video or pose, using methods such as CTC pretraining, temporal alignment, multimodal fusion, and language-model decoding. Researchers evaluate translation on benchmarks such as PHOENIX14T, CSL-Daily, How2Sign, and OpenASL, while also exploring low-resource languages and real-time streaming. A key tension is that standard text-overlap scores can miss meaning errors, and results may be inflated by signer overlap or inconsistent evaluation; reliable progress also depends on broader data and community-informed testing.
Latest papers
50 papers7 papers in the last four weeks, up 17% on the four weeks before. 0.0% of all new papers.