Google DeepMind Launches SL2T, a Sign-Language-to-Text Model Trained on 100,000+ Hours Across 50+ Languages

Google DeepMind has released SL2T, a sign-language-to-text translation model the company describes as a breakthrough in quality and generality for sign language AI. The model is now live in consumer products for the first time, powering sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with American Sign Language (ASL) to English.

What SL2T does

SL2T lets Deaf and hard of hearing users sign to their phone anywhere they would normally type — to search the web, draft messages, or issue queries to Gemini. In Live Transcribe, users can sign responses directly instead of typing back and forth during conversations. According to Google DeepMind's internal testers, signing in ASL was faster and more natural than typing in English for equivalent tasks.

Training data and architecture

SL2T is trained on more than 100,000 hours of data spanning over 50 sign languages, with roughly a quarter of that data in ASL. According to Google DeepMind, training jointly across multiple sign languages, dialects, and signer proficiency levels caused the model to learn shared underlying structures, outperforming single-language models in the company's internal experiments.

The model does not process raw video. Instead, an on-device model called MediaPipe Holistic tracks pose landmark locations — geometric coordinates of the signer's body, hands, and face — and only these coordinates are sent to the server for translation. The original video is discarded immediately, which Google DeepMind frames as a privacy safeguard.

SL2T translates the landmark coordinate sequence directly into text, skipping the intermediate "gloss" annotations used in prior sign language translation research. Google DeepMind argues that glosses fail to capture non-manual markers (facial expressions, head movement) and spatial grammar constructions that are integral to sign languages, and that direct landmark-to-text translation removes artificial vocabulary limits, allowing quality to scale with more data.

Benchmark results

On FLEURS-ASL (sd-test), a benchmark for ASL-to-English translation quality, SL2T achieves a zero-shot score of 70 BLEURT — which Google DeepMind describes as significantly higher than any previously reported score on this benchmark. No comparison scores from competing models were published alongside this claim.

Beyond the benchmark, Google DeepMind says it addressed real-world deployment issues including streaming latency, hallucination on non-signing inputs, fairness for left-handed signers (roughly 10% of the signing population), and accuracy for one-handed signing — relevant since users often hold the phone in one hand while signing with the other.

Rollout

SL2T is currently limited to ASL-to-English translation on Pixel 11 devices through Gboard and Live Transcribe. Google DeepMind says additional devices and sign languages are planned, though no specific timeline or list of upcoming languages was disclosed. Pricing is not applicable, as the feature ships as part of the Pixel software stack rather than as a standalone API product.

What this means

This is one of the first large-scale sign language translation models to reach mainstream consumer hardware rather than staying confined to research demos. The technical approach — training directly on pose landmarks instead of glosses, and pooling data across 50+ sign languages instead of building single-language models — is a meaningful departure from prior sign language AI research, which has historically been data-starved and fragmented by language.

The claims are notable but unverified by independent benchmarks: the 70 BLEURT score is Google DeepMind's own reported result on its own benchmark, and there is no third-party comparison yet. The real test will be how SL2T performs across signing styles, skin tones, lighting conditions, and non-native signers once it reaches a broader user base beyond internal testing. Limiting the initial release to a single device (Pixel 11) and a single language pair (ASL to English) suggests Google DeepMind is treating this as a controlled rollout rather than a finished product — expansion to more sign languages and devices will be the real indicator of whether this approach generalizes.