Roblox built an AI model that it says translates text chats so quickly users may not even notice it’s translating the messages of other players at first. It works with 16 languages, including English, French, Japanese, Thai, Polish, and Vietnamese. 

Dan Sturman, Roblox CTO, said in an interview with The Verge that the goal is to make Roblox users feel more comfortable engaging with each other by letting them understand what they are saying. The translator automatically translates chats, but users can click an icon to see the original message.

“We know engagement goes up when users speak or interact with others in their own language,” Sturman said. “We took that concept and removed the language barrier with the automatic translation.” 

Roblox began with building a transformer-based large language model (LLM), which trained on publicly available and internal data. It placed the LLM within a mixture of experts (MoE) architecture, an environment that ran multiple translation apps, with each one an expert in one language. This allowed Roblox to save on resources without needing to build a separate LLM for each language. Sturman said considering the scale of their project, his team believed it was easier to build their own model than modify an off-the-shelf LLM.

Sturman said Roblox monitors chats (it does so for trust and safety reasons) and can get feedback if translations are not exactly perfect. He added that the addition of the chat translation AI “changes nothing in our privacy and safety process, and banned words will still get blocked.”

Roblox, which has been trying to appeal to older audiences in the past few years, has been working with generative AI models to enhance user experience. It came out with an AI chatbot assistant for developers last year. It also automatically translates image assets, like words on buildings, to the user’s default language.

Sturman said the hope is for the translator model to move past just translating text chats eventually. “In the future, we could use AI to translate non-compliant [banned] words to compliant words or throw it at voice chats for real-time voice translation. There’s many possibilities,” Sturman said.

Other companies are also developing AI translator models. Meta released its speech-to-text and text-to-text translator, SeamlessM4T, which handles nearly 100 languages. Google’s Universal Speech Model also translates around 100 languages and is already deployed on YouTube to translate captions.