Roblox built an artificial intelligence model that can translate text chats so quickly that users won't even notice it's translating other players' messages at first. It is available in 16 languages, including English, French, Japanese, Thai, Polish and Vietnamese.

Roblox Chief Technology Officer Dan Sturman said the goal of the translator is to make it easier for Roblox users to understand what each other is saying, making them feel more comfortable communicating with each other. The translator automatically translates the chat, but users can click on the icon to view the original message.

"We know that engagement increases when users speak or interact with others in their own language," Sturman said. "We took this idea and removed the language barrier with automatic translation."

Roblox first built a transformer-based large language model (LLM), which was trained using public data and internal data. It places LLM in a Mix of Experts (MoE) architecture, an environment that runs multiple translation applications, each being an expert in one language. Sturman said that given the scale of their project, his team decided it would be easier to build their own model rather than modify an off-the-shelf LLM.

Stuurman said Roblox monitors chats (for trust and safety reasons) and can get feedback if the translation isn't perfect. He added that adding chat translation AI "will not change our privacy and security processes, and prohibited words will still be blocked."

Roblox has struggled to attract older users over the past few years and has been using generative artificial intelligence models to enhance the user experience. Last year, it launched an AI chatbot assistant for developers that can also automatically translate image assets, such as text on buildings, into the user's default language.

Stuhlman hopes the translation model will eventually move beyond mere text chat translation. "In the future, we can use AI to translate non-compliant [banned] words into compliant words, or use it for real-time voice translation for voice chats, there are many possibilities."

Other companies are also developing AI translation models. Meta has released SeamlessM4T, a speech-to-text and text-to-text translator that can handle nearly 100 languages. Google's universal speech model can also translate about 100 languages ​​and has been deployed to translate subtitles on YouTube.