On August 12, Google’s DeepMind unveiled SL2T, a brand-new multilingual sign language-to-text model. This sign language AI tech is now officially baked into the Pixel 11 phone system, marking the first time sign language AI has made its way into everyday consumer electronics—and finally bridging the smart device interaction gap for the deaf community.
The SL2T model, rolling out with the Pixel 11, initially pairs with two built-in tools: the Gboard keyboard and the real-time transcription app Live Transcribe. Once the front camera captures the user’s motions, they can edit messages, search the web, or chat with the Gemini large model—all by simply signing instead of typing manually. Internal beta data shows that users signing in American Sign Language can input text with better efficiency and more natural flow than traditional keyboard typing, dramatically improving the overall communication experience.
On the training front, SL2T was built on over 100,000 hours of sign language footage spanning more than 50 different sign languages, with a quarter of that data coming from American Sign Language datasets. This cross-lingual, joint training approach lets the model pick up on the common movement patterns shared across different sign languages, delivering better recognition than single-language models. It even set a new record on the FLEURS-ASL benchmark, raising the bar for current sign language transcription systems.
For years, voice-to-text and voice input have become standard features on smartphones, letting everyday users dictate text hands-free. But with over 200 sign languages in use globally and roughly 70 million deaf people, there’s long been a glaring gap: stable, mobile-friendly sign language recognition tools are practically nonexistent. Most sign language tech has been stuck in the lab, rarely reaching real-world users—leaving mobile accessibility for the deaf community sorely lacking.
This new feature gives deaf users a flexible “typing alternative” experience similar to voice input. Unlike older sign language recognition products that forced gestures to match fixed label vocabularies before converting to text, SL2T skips that clunky annotation step entirely. It directly parses body movement coordinates to generate text. Sign language is a full-fledged language in its own right—facial expressions, body positioning in space, and lip movements all carry grammatical meaning. Traditional label-based recognition tends to lose the semantic weight of non-hand movements, but this new architecture breaks free from the limits of fixed word banks, and its recognition ceiling keeps climbing as training data expands.
Privacy protection was also a top priority in this model’s design. The on-device MediaPipe Holistic tool tracks 130 key points across the hands, face, and torso in real time. The device only transmits movement coordinate data—raw video footage is deleted instantly and never uploaded to the cloud, eliminating the privacy risks of leaked camera captures.
According to the DeepMind team, the project brought deaf professionals and a sign language advisory board into the product iteration loop from day one, ensuring the technology matches how signers actually communicate day-to-day. For now, the feature only supports American Sign Language to English, but Google plans to gradually expand to more sign languages and roll the accessibility feature out to more Android devices down the line.
This isn’t DeepMind’s first foray into accessibility AI for people with disabilities. Back in May 2025, the research team open-sourced SignGemma, a sign language translation model built on the lightweight Gemma architecture. That project, primarily optimized for American Sign Language, laid the algorithmic groundwork for SL2T’s consumer-ready debut.
The lab has also developed a range of other inclusive tech: using its proprietary WaveNet speech synthesis algorithm to help ALS patients recreate their original voices; running the Euphonia project to recognize slurred or unclear speech from stroke and aphasia patients, breaking down device interaction barriers for those with speech impairments. On the mobile side, Google’s accessibility product line Live Transcribe has spent years refining audio transcription algorithms, providing daily captioning for deaf users. That years of acoustic-visual multimodal experience has all been poured into this sign language recognition model.
Several tech companies are already racing to build AI accessibility solutions for people with disabilities. In China, iFlytek’s Tingjian (iFlytek Hearing) offers free real-time voice transcription and monthly-quota conference interpretation for users certified with hearing disabilities. Huawei, powered by its HarmonyOS large model, has built a comprehensive assistive toolkit—its AI glasses use the “Xiaoyi Sees the World” feature to help visually impaired users detect road obstacles, and Xiaoyi AI voice restoration improves unclear speech for those with speech loss. Apple has upgraded on-device AI for iPhone and Vision Pro, equipping devices with intelligent real-time captions and image detail description, while Vision Pro adds a new eye-tracking accessibility feature that even works with external powered wheelchairs.
