Rylo, a company with Tel Aviv and New York hubs, announced an NVIDIA collaboration on October 21, 2026 to advance research into AI-powered sign-language translation.
CTech’s October 21 report independently covered the collaboration and an accompanying desktop captioning product. The announcement brings together a longer-term research effort and a practical question for workplace software: how to make conversations easier to follow and participate in.
A direct connection to Tel Aviv and New York
Rylo’s careers page identifies hubs in both cities and lists Tel Aviv roles in frontend and backend engineering. The company, formerly known as Nagish, is therefore relevant to TLVNY through its operating team, alongside the accessibility products it is developing.
For readers following the two technology communities, the story offers a concrete example of an AI application outside general-purpose chat: software designed around communication access. Its progress will depend on whether people can use it reliably in everyday conversations.
Why video representation matters
In the dated research announcement, research head Dr. Amit Moryossef says NVIDIA is providing supercomputer access. The team is adapting Cosmos-based video tokenization for models that process and generate signing video, using NeMo Megatron Bridge for training.
NVIDIA’s Cosmos Tokenizer explanation describes the underlying idea: an encoder converts images or video into a smaller representation, and a decoder reconstructs visual content from it. The company provides continuous and discrete tokenizers. Discrete tokens are numerical identifiers that fit the next-token prediction approach used by GPT-style models; continuous representations store vectors of numbers. Both involve a trade-off between compactness and retained visual information.
This background explains the engineering problem without establishing that a particular sign-language application has solved it. A smaller representation is useful for computation, but the communication value depends on what survives the transformation.
NVIDIA’s Megatron Bridge documentation describes a training library supporting pretraining and fine-tuning of language and multimodal models, with tools for model conversion and parallel training. That places the collaboration within model development infrastructure. The presence of a training framework is not, by itself, a measure of translation quality.
Moryossef’s research account acknowledges remaining problems with hands and other fine details, and says native signers are helping evaluate intelligibility. It describes progress toward automatic interpretation, rather than announcing a finished conversational translation service.
Building with people who sign
Rylo’s separate sign-language project page describes a team of Deaf and hearing researchers, engineers, linguists and designers. It explains that its modular approach separates visual understanding from language translation, recognizing the importance of handshapes, facial expressions, spatial features and grammar. The page also presents existing Sign Translate and dictionary tools; these should be distinguished from the newly described NVIDIA collaboration.
From TLVNY’s perspective, evaluation should follow the conversation rather than stop at a visually convincing demonstration. A useful assessment would ask whether signers understand the intended meaning, how errors are identified and corrected, and what happens when a system is uncertain. Those are editorial evaluation questions, not results reported for this project.
A separate desktop product for live captions
For people interested in a tool they can investigate now, Rylo Desktop’s current product page, checked on October 22, lists Mac and Windows support and provides a download link. Rylo describes it as free, without a subscription or per-minute charge, and says it works alongside meeting software including Zoom, Google Meet and Microsoft Teams.
The page describes several caption layouts, speaker identification and a Type-to-Speak function that reads typed replies aloud. It also advertises searchable transcripts and AI summaries. Rylo says no separate bot joins the meeting and transcripts remain on the user’s device. These are the company’s product descriptions, rather than findings from a TLVNY installation or security test.
Its FAQ says live captions require an internet connection, even though saved conversations can be accessed offline. It lists 60 caption languages. That spoken-conversation captioning capability is separate from the sign-language research; it should not be read as a claim that the new research translates 60 signed languages.
What would make the next update meaningful
The next useful evidence would connect technical development to communication outcomes: clearly defined tasks, evaluation by signers, documented limitations and performance across realistic settings. For workplace users, equally practical questions concern caption accuracy during overlapping speech, control over saved transcripts and how a typed response enters a fast-moving meeting. Those questions give readers a way to assess future updates while keeping today’s announcement in proportion.
