手話AIをユーザーの手に：SL2Tモデルの発表Putting sign language AI into users’ hands

Google DeepMind has introduced SL2T, short for sign-language-to-text, a model the company describes as a breakthrough for translating signed communication into written text. The announcement matters because it aims to bring practical sign language recognition directly to Deaf and hard of hearing users, a group that has historically been underserved by mainstream speech and language technology.

At its core, SL2T is designed to take visual input of a person signing and produce corresponding text. According to Google DeepMind, the model powers new sign language features intended to make everyday technology more accessible. Rather than positioning the work purely as a research demonstration, the framing of putting sign language AI in users' hands suggests the company intends the capability to reach real users through consumer-facing features rather than remaining confined to a laboratory setting.

Sign language recognition is a substantially different problem from speech recognition, even though the two are often grouped together as accessibility technologies. Signed languages are full natural languages with their own grammar and vocabulary, and they convey meaning through a combination of hand shape, movement, location, facial expression, and body posture. This multi-channel, spatial nature makes them difficult to capture with models built primarily for text or audio. A system must interpret continuous motion, handle variation between signers, and account for the fact that sign languages are not universal, as American Sign Language, British Sign Language, and many others differ significantly.

One persistent obstacle in this field has been the scarcity of large, high-quality annotated datasets compared with the enormous text and audio corpora available for training language and speech models. Progress in sign language processing has therefore often lagged behind advances in automatic speech recognition and machine translation. A model that meaningfully improves recognition accuracy and can be deployed to users would represent notable progress, though independent evaluation would be needed to confirm how well it performs across different signers, sign languages, and real-world conditions.

The release fits within a broader pattern of accessibility work at Google. The company has previously shipped tools such as Live Transcribe and Live Caption, which turn spoken audio into on-screen text in real time, along with Sound Amplifier for people with hearing difficulties. Project Euphonia focused on improving speech recognition for people with atypical speech patterns. SL2T appears to extend this line of work from audio-based accessibility toward visual, gesture-based communication, addressing a modality that earlier tools did not cover.

The model is presented under Google DeepMind, the research organization behind the company's Gemini family of models, and its categorization alongside Gemini suggests it draws on the group's wider expertise in multimodal AI, meaning systems that process images, video, and language together. Multimodal understanding is a prerequisite for interpreting sign language, since the task requires linking visual sequences to linguistic output. The available announcement does not, however, detail the specific architecture, training data, supported sign languages, or the platforms and regions where the features will be available.

Broader industry interest in sign language technology has grown in recent years, with academic groups and startups exploring recognition and translation systems, and organizations within Deaf communities emphasizing that such tools should be developed with, and not merely for, the people who use them. Sign language advocates have frequently cautioned against overstating what automated systems can do, noting that nuances of expression and regional variation remain challenging. Google DeepMind's framing of the model as a step toward accessibility, rather than a complete replacement for human interpreters, is consistent with that caution.

For Deaf and hard of hearing users, the practical value will depend on how the features are integrated into products and how reliably they work outside controlled demonstrations. As with earlier accessibility launches, adoption is likely to hinge on accuracy, latency, privacy protections for camera-based input, and the range of sign languages supported. Further technical documentation and independent testing would help clarify the model's real-world capabilities and limits as it moves from announcement to deployment.