Google's New Pixel Can Read Your Sign Language and Turn It Into Text

Google launched its new Pixel 11 phone on August 12, 2026, and it includes a feature that lets deaf and hard of hearing people sign to their phone and have those signs appear as written text in real time (Engadget).

The feature comes from a new model called SL2T, built by Google's DeepMind team. It is embedded directly into Gboard, the keyboard app that comes default on Android phones, and Live Transcribe, Google's existing tool that turns spoken words into on-screen captions. That means a person can sign to their phone wherever they would normally type.

At launch, SL2T only works for American Sign Language (ASL) converted into English. That is a narrow starting point. But DeepMind trained the model on more than 100,000 hours of sign language footage in multiple languages, with about a quarter of that in ASL. The company says training on several sign languages at once helped the model learn patterns that are shared across them, which could make it faster to add more languages later.

The design also protects privacy. The phone does not send video of you to Google's servers. Instead, a model running on the phone itself turns the camera image into a kind of stick-figure skeleton — a set of coordinates that tracks the movement of your hands, arms, and body. Only that abstract skeleton data gets sent to Google's servers for transcription. The server never sees your face or what you look like. Sending skeleton data instead of full video also uses less internet bandwidth.

SL2T also skips a step that most sign language recognition systems have used. Older systems typically first label each individual sign with a word tag — researchers call these tags "glosses" — and then translate those tags into the target language. SL2T goes straight from the skeleton data to finished text, cutting out the gloss step entirely. This follows a trend in other areas of artificial intelligence, where researchers have been removing intermediate steps in favor of systems that go directly from input to output.

DeepMind announced the feature in a blog post titled "Putting sign language AI into users' hands" (DeepMind blog). The post came out alongside the Pixel 11 hardware launch.

The potential audience is large. More than 70 million deaf and hard of hearing people around the world use roughly 200 different sign languages. ASL-to-English covers just one of those 200 language pairs. Google has not announced a timeline for adding more languages beyond what shipped at launch.

Putting the feature inside Gboard and Live Transcribe was a deliberate choice. These are tools people already use every day — one for typing, one for reading captions of spoken words. By building sign language transcription into those familiar places, the feature reaches people where they already are, rather than requiring them to find and download a separate accessibility app. Accessibility tools that live inside their own dedicated apps tend to reach only the people who already know to go looking for them.

The broader context here is something we have watched happen across the smartphone era. A feature that once needed special hardware or its own separate software gradually moves into the phone's built-in operating system. Live Transcribe itself followed this path when it first launched for speech. SL2T applies the same idea to sign language.

In my view, the split between phone and server is the most interesting design decision. The phone handles the privacy-sensitive part — turning your image into an anonymous skeleton — while the heavier work of actually reading the signs happens on Google's servers. Doing everything on the phone would likely make it slow and hot, which would hurt the real-time typing experience. Doing everything on the server would mean sending video of yourself. The skeleton approach is the middle ground that keeps both privacy and speed acceptable.

For developers working on accessibility, the decision to skip the gloss step is the detail worth watching. If SL2T's direct approach works well in everyday use, it could change how other teams build sign language tools, especially for languages where creating a full gloss dictionary would be too difficult or expensive. The multilingual training helps here too: adding a new language should require less data than starting from zero.

The Pixel 11 puts this feature into users' hands on day one. Whether it grows beyond a single language pair, and how fast, will determine whether SL2T becomes a genuinely useful tool for the global deaf community or stays a promising but narrow demo of what the technology can do.