From Clips to Streams: A Unified Framework for Streaming Sign Language Translation

Abstract

Sign Language Translation (SLT) research has typically focused on pre-segmented clips, a constraint that disconnects models from the continuous reality of real-world communication. To bridge the gap from clips to streams, we formally define Streaming SLT, a realistic task demanding the joint discovery and translation of linguistic events from continuous, untrimmed video inputs. To address this challenge, we propose StreamSLST, the first end-to-end framework for Streaming Sign Language Segmentation and Translation via Deformable Transformer architecture. Unlike cascaded approaches that suffer from error propagation, StreamSLST employs a parallel decoding strategy to jointly optimize temporal segmentation and translation in a single pass. To enable scalable training on long-form videos, we incorporate a Sentence-aware Sliding Window Sampler that crops continuous streams into trainable units. We further design a Tri-modal Visual-Language Pretraining stage that aligns pose dynamics with textual semantics before joint optimization. To support comprehensive evaluation, we conduct experiments on two real-world streaming datasets, BOBSL and How2Sign, together with two newly constructed synthetic streaming benchmarks, Streaming-Phoenix-2014T and Streaming-CSL-Daily. The results demonstrate that StreamSLST consistently outperforms a matched cascaded segmentation-then-translation pipeline. Interestingly, our analysis reveals that joint optimization learns wider, translation-effective temporal support that simple expansion or rescaling of the annotated spans fails to reproduce. Our work establishes the first robust benchmark for continuous sign language understanding and paves the way for accessible real-world communication interfaces. The source code and data are available at https://github.com/streamsl/StreamSLST.