AI sign segmentation eases lookup in a 13-learner ASL study

Thirteen hearing, intermediate American Sign Language learners reported lower workload when a video lookup tool automatically split footage into signs. The comparison tool required users to select a video segment manually. Both tools used AI recognition: the experiment tested the additional segmentation step, not AI against a wholly manual dictionary.[1]

Each learner tried both tools using four selected clips on a laptop. Search satisfaction rose by 0.5 points on a five-point scale. Translation scores did not differ significantly; that result does not establish that the tools perform equally well.[1]

The workload measure, NASA-TLX, asks users to rate mental and physical demands, time pressure, effort, frustration and perceived performance. It measures the experience of doing a task, rather than independently checking whether a translation is correct. Here it helps separate a useful interface improvement from a claim of better language understanding.[3]

The paper treats its analyses as exploratory. The 26 observations per condition came from repeated measurements of 13 people, and that dependence was not fully accounted for. Model training also included the study clips, limiting what the results establish about unfamiliar footage.[1]

Published at ASSETS 2026, the study offers a promising workflow for intermediate learners. Its research code is available, but requires technical setup; the results do not establish performance for Deaf native signers or an unrestricted video service.[1][2]

Report an error or suggest a correction