Online gaming giant Roblox has unveiled an innovative AI-powered “unified translation model,” enabling players to communicate in real-time through text, even if they speak different languages.
|Chinese – Simplified
|Chinese – Traditional
Sturman emphasized the importance of creating a translator that goes beyond translating static content within gaming experiences. The key challenge was developing a system that independently translates between all 16 languages while ensuring real-time capabilities, achieving a base latency of 100 milliseconds for instant conversations.
How Roblox built it
The process involved Roblox building its own large language model (LLM) using a transformer-based approach. Rather than creating unique models for each language pair, the company opted for a unified LLM, capable of handling all language pairs in a single model.
To train the model, Roblox used a combination of public and private data and engaged expert translation apps for each language. Overcoming challenges in less common language pairs, such as French to Thai, required the use of “back translation” to ensure accuracy.
Human evaluators played a crucial role in training the system to understand slang and incorporate popular and trending terms in multiple languages. This ongoing process ensures that the AI translator stays updated with the latest phrases and expressions.
In the final-stage testing, Roblox observed that the new translation system significantly enhanced user engagement and session quality on its platform. With 70 million daily active users across 180 countries exchanging over 2.4 billion messages daily, the real-time AI translator is expected to elevate the gaming experience for a diverse global audience.