Suvana: An Integrated Platform for Two-Way Deaf–Hearing Communication in Sri Lankan Sign Language

Abstract Sri Lankan Sign Language (SSL) is documented as under-served: interpreter provision is limited, no fully stan-dardised national variety exists, and sign-language technology overwhelmingly targets American Sign Language. We present Suvana , a platform that brings four SSL components under one deployment: isolated sign recognition with Sinhala speech output, Sinhala speech to a signing avatar, on-device ambient-sound alert-ing, and an in-browser practice tutor for hearing learners. Our contribution is twofold. First, we describe an integration topology that lets four independently developed research prototypes — three web applications, one mobile application and two Python services — be served as a single product without merging their codebases, including a documented exception for the component whose transport cannot be proxied. Second, we report a cali-brated Dynamic Time Warping (DTW) grader for the practice module and give measured evidence that open-set identification on the same distance metric would be unsafe: the closest pair of distinct SSL signs in our reference library sits 0.135 apart in normalised hand-size units, while two correct renditions of the same sign sit a median 0.458 apart. We therefore grade a specified sign rather than classify an unknown one. We report the practice module’s discrimination as ROC AUC 0.744 against a 65.4% majority-class baseline, and state explicitly which of our targets remain unmeasured. We also report the recognition module’s accuracy under the signer-dependent protocol it was trained with, and decline to interpret it as generalisation performance. All four authors are hearing; no Deaf signer has yet validated any component, and we set out what that means for the claims we make.