Google DeepMind introduced SL2T today: a massively multilingual sign-language-to-text model shipping immediately in Gboard and Live Transcribe on Pixel 11, American Sign Language to English, free. Of the world's 200-plus sign languages and an estimated 70 million Deaf and hard-of-hearing signers, this covers exactly one language pair on one device — worth stating plainly before the rest of what's genuinely well-built here.
The actual technical bet: skip the gloss
Most prior sign-language translation work routes through glosses — an intermediate word-by-word annotation layer standing in for the signed sequence before a separate step turns that into fluent target-language text. Google's argument is that glosses are the wrong abstraction: they're built for spoken-language word order and can't represent non-manual markers (facial expression, mouth movements) or spatial constructions (where signs reference locations in the space around the signer) that carry real grammatical meaning in sign languages. SL2T translates directly from pose-landmark coordinates to text, no gloss step, which the team says removes an artificial vocabulary bottleneck and lets translation quality scale directly with training data rather than with how complete a hand-built gloss dictionary is.
That's the same category of problem DeepMind's AMIE project hit from a different angle — perception and reasoning that have to work over continuous physical movement, not discrete tokens — solved here by refusing an intermediate representation rather than building a better one.
A genuinely specific privacy architecture
SL2T never sees a face or a body on the server. An on-device model (MediaPipe Holistic) extracts pose landmark coordinates — points on the signer's hands, arms, torso, head — locally, and only those geometric coordinates are transmitted for translation; the source video is discarded immediately. That's a concrete, verifiable architectural claim rather than a policy promise, and it matters specifically for a signing population, since a video feed of someone's face and hands is far more identifying than typed text ever was.
The numbers, and the one that's missing one
Training data: 100,000+ hours across 50+ sign languages, roughly a quarter of it ASL, trained jointly rather than as separate single-language models — Google reports the multilingual approach outperforms single-language training, evidence of real cross-lingual transfer. On FLEURS-ASL (sd-test), an ASL-to-English benchmark, SL2T scores 70 BLEURT zero-shot, which Google calls "significantly higher than any previously reported score" — a claim we can't fully evaluate without the specific prior number, though FLEURS-ASL is an independent academic benchmark rather than a vendor-chosen comparison set, which is a meaningfully different credibility position than the self-selected comparison tables this month's model launches have leaned on.
The error table is the part worth crediting
Google published five real FLEURS-ASL translation examples with source and output side by side, and named the specific failure modes still present: rare signs, fingerspelling errors under rapid signing ("prey" mistranslated as "grey"), dropped classifier depictions (a description loses "claws" entirely), passive-construction handling, and tense ambiguity without surrounding context ("kicked off" becomes "start"). That's a genuinely useful disclosure — a labeled taxonomy of how the system fails, not just a claim that it works — in the same spirit we credited Ling-3.0-tiny's card for disclosing its harness and AMIE's writeup for stating its study was entirely simulated. Two engineering details worth noting for the same reason: the team specifically tuned for the ~10% of signers who are left-handed, and for one-handed signing — the realistic case of someone holding a phone in their other hand — rather than optimizing only for a clean, two-handed lab benchmark.
The governance structure is unusual for this industry
SL2T was conceptualized by a Deaf Googler (Sam Sepah), developed with Deaf data-collection partners, evaluated in Deaf user studies, and is now overseen by an AI Sign Language Advisory Committee (AISLAC) convening global Deaf organizations — with a joint impact report co-authored with that committee documenting the system's capabilities and limitations at launch, a practice Google says will continue for future sign-language releases. Participatory governance where the affected community has standing input before and after release is not the norm for AI launches this blog covers; it's worth naming as a real difference in process, not just a line in a press release, precisely because the error-table disclosure above suggests the commitment produced an actual artifact rather than a talking point.
What to expect next
- Watch the "more languages coming" timeline specifically. The 50-language training corpus is a research and transfer-learning asset; only ASL ships today. The gap between "trained on" and "available in" is the thing to track, the same gap this month's open-weights sagas kept surfacing in a different domain.
- Watch for sign-language generation, the reverse direction (text/speech to sign), which Google names explicitly as future work and which is a substantially harder generation problem than translation into text.
- Watch whether AISLAC's model — impact report plus named advisory committee — becomes something other AI accessibility efforts adopt, or stays specific to this one team's practice.
References: Google DeepMind — Putting sign language AI into users' hands · Google DeepMind on X · related coverage: AMIE (Video) · Ling-3.0-tiny · Gemini 3.7 Flash