Roblox builds in-house LLM, adds real-time AI translation to metaverse

Online gaming platform Roblox has developed a new AI-powered “unified translation model” — allowing players to converse with one another through text in real-time, despite speaking different languages.

According to a Feb. 5 statement by Roblox chief technology officer Dan Sturman, the new system required the firm to build its own in-house LLM to be able to translate text-based messages at a base latency of 100 milliseconds — making conversations feel instantaneous for users.

“Imagine discovering that your new Roblox friend, a person you’ve been chatting and joking with in a new experience, is actually in Korea — and has been typing in Korean the entire time, while you’ve been typing in English, without either of you noticing,” said Sturman.

The LLM translates text-based messages in real-time across 16 supported languages. Source: Roblox

How Roblox built the “unified translation model”

While Sturman noted that Roblox already automatically translates its in-experience content, the firm wanted to “go beyond translating static content in experiences.”

Sturman said that the two largest obstacles to creating the translator were designing a system that allowed for translation between all 16 languages independently and then making it fast enough to support real-time chats, which required a novel approach to building its own large language model (LLM).

“To achieve this, we could have built out a unique model for each language pair (i.e., Japanese and Spanish), but that would have required 16×16, or 256 different models. Instead, we built a unified, transformer-based translation LLM to handle all language pairs in a single model.”

Roblox’s AI translator began with the firm’s creation of a transformer-based LLM trained on public and private data.

Roblox then handed the LLM over to a mixture of “expert” translation apps, which trained the model on each individual language.

Sturman noted that “less common” translation pairs, such as French to Thai, were challenging due to a lack of high quality data, forcing Roblox to use “back translation” where messages are translated into the original language and then compared with source text for accuracy.

Following this, the translations are parsed through a quality estimation model which focuses primarily on the understandability of the translation.

The Roblox translator parses phrases through a quality estimation model. Source: Roblox

Additionally, the model was trained to understand human slang, with Roblox bringing in human evaluators to translate “popular and trending terms” for each language.

Sturman added that this was an ongoing process, with human evaluators constantly updating the system to keep it up to speed with the latest phrases.

In final-stage testing, Roblox found that the new translation system drove “stronger engagement and session quality” for users of its platform.

Roblox currently boasts 70 million daily active users from more than 180 countries around the world, with over 2.4 billion messages being exchange every day.

In November last year, Roblox CEO David Baszucki said he “dreams” of interoperability and believes all metaverse users should be able to move nonfungible tokens (NFTs) and other digital assets between multiple independent platforms.

Main, News