Since August 12, 2026, Google has offered a feature on the Pixel 11 that can translate American Sign Language into English text in real time within Gboard and Live Transcribe. The SL2T model combines computer vision and machine translation, processing video locally and then running computation on servers using body coordinates. This first consumer-facing integration broadens the available input methods, but it does not yet cover all Android phones or other sign languages, and it is not designed to replace an interpreter in sensitive situations.
What Google actually launched
Google DeepMind and Android have integrated Sign-Language-to-Text, or SL2T 1.0, into Gboard and Live Transcribe. The feature converts signs into text as the signer moves. In Gboard, this text can feed a messaging app, a note, a search or a request to Gemini. In Live Transcribe, the goal is to make informal face-to-face conversation easier.[1][2]
The initial scope is narrow. It requires a Pixel 11, the latest version of Gboard and a keyboard set to US or Canadian English. The only language pair announced at launch is American Sign Language to English. Google plans to add other devices and languages, without a specific public timeline.[1][2]
Describing this as a general translation of “sign language” on Android would therefore be inaccurate. Sign languages are distinct natural languages, each with its own grammar and lexicon. French Sign Language, for example, is not a visual variant of French and is not supported in this version.
Why this translation is more complex than speech transcription
Speech recognition generally maps an audio signal to text in the same language. SL2T has to translate between two languages. Meaning is carried simultaneously by the hands, arms, torso, head, face, space and rhythm. An approach that converted each sign into an English word would miss part of this structure.
The system therefore translates a sequence of movements directly into text, without going through “glosses,” the intermediate annotations often used in research. This choice may better preserve certain spatial and non-manual constructions, but it also makes error analysis harder: the model produces a likely sentence without providing a verifiable intermediate linguistic transcription.[1][3]
How SL2T turns gestures into text
The front-facing camera observes the face, upper body and hands. MediaPipe Holistic, running on the phone, extracts 130 two-dimensional landmarks from each frame. The raw video is then deleted. Only the sequence of coordinates is sent to Google’s servers, where a Transformer-type network generates the English translation.[2][3]
This representation greatly reduces the amount of information transmitted and conceals part of the person’s appearance and surroundings. It is not neutral, however: it loses depth, certain facial details and articulators such as the tongue, as well as objects a person may refer to. The AISLAC joint report stresses that these losses can affect comprehension.[3]
What the results actually support
Google says it trained SL2T on more than 100,000 hours of data covering more than 50 sign languages, about a quarter of it in ASL. The lab claims this multilingual training improves transfer across languages and variants. However, the model weights, the full data and the training protocol have not been published, which limits independent replication.[1]
On FLEURS-ASL, a studio-recorded ASL-to-English translation dataset, SL2T achieves a BLEURT score of 70 in zero-shot evaluation and a BLEU score of 25. The report also cites a BLEURT of 74 on a one-handed signing subset. BLEURT measures the semantic closeness between a translation and a reference, but it does not directly reflect how safe a real conversation is.[1][3]
The published examples still show confusion on rare signs, fast fingerspelling, verb tense and certain spatial descriptions. No large-scale independent evaluation was available as of September 23, 2026. The launch therefore demonstrates a useful capability within a defined scope, not universal translation or equivalence with a professional interpreter.
An accessibility advance, provided users keep their choice
The most concrete benefit is a new input method. A signer can produce text without constantly relying on an English keyboard, even though written English may be a second language. The tool can reduce everyday friction in messages, searches and short exchanges.
This usefulness should not lead to machine translation being imposed as the default interface. Deaf and hard-of-hearing people have diverse language practices. Some prefer a sign language, others writing, speech, captioning or a combination. Accessibility therefore requires multiple options, easy correction and the ability to switch back to the keyboard.
Participatory governance is a notable element here. Google created the AI Sign Language Advisory Committee with associations, academic institutions, interpreters and Deaf experts. The joint report documents the intended uses and limitations. This consultation does not replace an independent audit, but it gives the communities concerned a more substantial role than a mere testing phase.[3]
Privacy: a protective architecture, but not fully local
The phone does not transmit raw video. This minimization reduces risks related to images of the face, the home and the surroundings. Google also states that coordinates, videos and translations are neither retained nor used for training, unless explicitly authorized as part of a study.[2][3]
Translation nonetheless remains cloud-dependent. Body coordinates are data derived from a person’s behavior and may fall under the GDPR when they relate to an identifiable user. They only become sensitive biometric data within the meaning of Article 9 if they are processed to uniquely identify a person. The legal basis, user information, transfer security and the effectiveness of non-retention therefore remain compliance issues to be documented.[4]
Accessibility and law: what the European framework implies
The European Accessibility Act has applied since June 28, 2025, to a number of products and services, including smartphones and certain digital services. It sets functional accessibility requirements without prescribing a technical solution such as SL2T. Such a feature can contribute to accessibility, but its presence is not enough to demonstrate a product’s overall compliance.[5]
The EU Artificial Intelligence Act does not automatically classify a consumer translation tool as a high-risk system. The context of use remains decisive. Using an unevaluated translation in a medical, legal, educational or professional decision would increase the consequences of an error and could trigger other sector-specific obligations. The explicit restriction to low-stakes situations is therefore essential.[3][6]
What to watch next
- The actual rollout of other sign languages, with data, evaluations and consultations specific to each community.
- Availability on other Android devices and real-world performance depending on the camera, lighting, body type and framing.
- Independent evaluations of accuracy, latency, serious errors and differences between groups of signers.
- The possibility of more on-device processing, to reduce dependence on the network and the exposure of body coordinates.
- Continued access to human interpreters when accuracy, confidentiality or accountability require it.
SL2T marks an important milestone because a sign language translation model is leaving the lab and becoming a consumer input tool. Its value, however, will not be measured by a benchmark score alone. It will depend on real-world reliability, expansion to more languages, data protection and, above all, the ability to strengthen signers’ autonomy without narrowing their choices.
Further reading
To explore the links between accessibility, multimodal interfaces and on-device AI processing, check out these analyses from the aivancity blog.
Sources
[1] Google DeepMind, August 12, 2026, Putting Sign Language AI into Users’ Hands. Read the article
[2] Google Pixel Help, Use Sign-to-Text to Translate American Sign Language into English Text in Real Time, accessed September 23, 2026. View the documentation
[3] Google DeepMind and AISLAC, August 12, 2026, AISLAC Joint Impact Report for SL2T 1.0. View the report
[4] European Union, Règlement (UE) 2016/679 relatif à la protection des données (RGPD), Articles 4 and 9. View the regulation
[5] European Commission, European Accessibility Act, official fact sheet. View the official fact sheet
[6] European Union, Règlement (UE) 2024/1689 établissant des règles harmonisées concernant l’intelligence artificielle (AI Act). View the regulation