Real-time artificial language translation added to the Roblox Metaverse by in-house LLM

With the help of a new AI-powered “unified translation model,” built by the online gaming platform Roblox, users may now communicate textually in real time even when they speak various languages.

In order to translate text-based messages at a base latency of 100 milliseconds, the new system needed Roblox to develop its own in-house large language model (LLM), according to a statement released by the company on February 5. This would let users to feel as though discussions are happening instantly.

Imagine finding out, without either of you realizing, that your new Roblox buddy-someone you’ve been laughing and conversing with in a novel experience-is actually in Korea and has been typing in Korean the entire time, while you’ve been typing in English, according to Sturman.

How the “unified translation model” was constructed by Roblox:

Roblox already translates its in-experience content automatically, as Sturman pointed out, but the company aimed to “go beyond translating static content in experiences.”

According to Sturman, the two biggest challenges in developing the translator were coming up with a system that could translate text independently across all 16 languages and then speeding it quickly enough to enable real-time conversations, which required a creative way of developing its own LLM.

The company’s transformer-based LLM, trained on both public and private data, served as the foundation for Roblox’s AI translator.

After then, Roblox gave the LLM to a variety of “expert” translation programs, which trained the model on data from each distinct language.

Roblox had to for to “back translation,” in which messages are translated into the original language and then checked for correctness against the source text, because Sturman pointed out that “less common” translation combinations, like French to Thai, were difficult because there wasn’t enough high-quality data.

After that, a quality estimation methodology is used to the translations, with a primary focus on the translation’s understandability.

Furthermore, Roblox employed human evaluators to interpret “popular and trending terms” for every language so that the model could be trained to comprehend slang from other languages.

According to Sturman, the process is continuous, since human evaluators continuously update the system to ensure it remains current with the most recent terms.

Roblox discovered that the new translation system increased “stronger engagement and session quality” for its platform users during the last round of testing.

At the moment, Roblox has 70 million active daily users from more than 180 countries, and over 2.4 billion messages are sent and received everyday.

CEO of Roblox David Baszucki stated in November 2023 that he hopes for interoperability and that all users of the metaverse should be able to transfer digital assets, such as nonfungible tokens, between various independent platforms.