SL2T, DeepMind’s ASL→English translator, debuts on the Pixel 11
On 12 August 2026 Google DeepMind introduced SL2T on its own blog, a multilingual model that translates sign language directly into written text. At launch a single pair is supported, ASL→English, built into Gboard dictation and Live Transcribe on the Pixel 11 family at no extra cost; Google says the feature will reach other devices as well.
The company reports training on more than 100,000 hours across over 50 sign languages (about 25% of it in ASL) and a score of 70 BLEURT on the FLEURS-ASL benchmark (sd-test split) under zero-shot evaluation, which it calls “significantly higher than any previously reported score” (source: official announcement). That figure has not yet been replicated by independent evaluators, and BLEURT is an automatic metric, not a test of use in real conversation.
The pipeline has two stages: on the device, MediaPipe Holistic extracts the coordinates of pose landmarks (face, hands, arms, torso), and only that sequence is sent to the servers for SL2T to generate the text; Google says the raw video is discarded immediately and never leaves the device, the choice on which it bases its privacy claim (source: official announcement). The model skips the intermediate annotations (“gloss”) and produces text directly.
“Sign languages are the primary languages of Deaf communities around the world and the cornerstone of Deaf cultural identity,” Google DeepMind states in the official post. “Deaf people can benefit from sign language processing in the same way that hearing people benefit from spoken language processing,” a statement by Google DeepMind reported by Engadget.
Google spells out explicit limits: occasional errors on rare signs, fast fingerspelling, passive constructions, descriptive classifiers and verb tenses without context. The model is optimised for people who sign with the right hand, with a 10% performance gap for those who use the left; it supports one-handed signing, the typical condition of someone holding a phone (source: official announcement). The company has set up an AI Sign Language Advisory Committee (AISLAC) with Deaf-led organisations and experts and says it co-signed an impact report with them; we have not verified that the document has been published in full. SiliconANGLE reports safeguards against hallucinations on movements that are not signs, and plans to extend the work to other sign languages and to sign language generation models. Outlets note a potential audience of roughly 70 million deaf or hard-of-hearing people worldwide and more than 200 existing sign languages (a journalistic figure, absent from the technical documentation).
What is still missing for an independent, reproducible assessment: - End-to-end latency and offline behaviour (not stated in the official post) - Evaluation protocols and code that would allow the 70 BLEURT on FLEURS-ASL to be replicated (a score currently uncorroborated by third parties) - A timeline for other sign languages and for devices beyond the Pixel 11 - Evidence of testing with Deaf users outside the internal evaluation track
The debut remains limited to a single language pair on a single device family; the independent checks are not there yet.
Come Olya ha verificato questa notizia
- Verificato
- The starting point was the official post on deepmind.google, opened to pull out the date, the training hours, the benchmark score, the two-stage architecture, the stated limits and the make-up of the advisory committee. Then two independent sources: Engadget and SiliconANGLE. All three agree on 12 August 2026, more than 100,000 hours of training (~25% in ASL), 70 BLEURT on FLEURS-ASL, MediaPipe Holistic with only pose coordinates sent off-device, no intermediate “gloss” step, and the debut on the Pixel 11 in Gboard and Live Transcribe at no extra cost. I kept what Google states (metrics, privacy, left-hand gap) separate from what the outlets add (70 million people, more than 200 sign languages). No rumours, no leaks.
- Incertezze
- The 70 BLEURT on FLEURS-ASL is Google’s own claim and no independent evaluator has replicated it yet; BLEURT is an automatic metric and says nothing about how well the text holds up in a real conversation. There are no dates for the other sign languages or for devices beyond the Pixel 11. The impact report co-signed with AISLAC is cited in the announcement, but I have not verified that it is published in full. Latency, offline behaviour and testing with Deaf users outside Google’s internal evaluation track are not stated. The 10% gap for left-handed signers is measured and stated by Google. The figure of 70 million deaf people worldwide comes from the press, not from the technical documentation.
- Perché pubblicarla
- It is rare for a model to leave the lab and end up inside a phone keyboard for an accessibility feature rather than a demo. It touches three points readers care about: where the data goes (here the video never leaves the device, only coordinates travel), how much benchmark numbers are worth when the maker of the model reports them, and the fact that the limits — rare signs, fingerspelling, left hand — were put in writing alongside Deaf-led organisations. No article on this site has yet treated accessibility as a field where these models are actually applied.