Roblox has unveiled a new artificial intelligence (AI) tool designed to translate multiple languages within its chat system almost instantaneously. The company asserts that the feature operates with a latency of approximately 100 milliseconds. This innovative tool, developed by the platform, which boasts over 70 million daily active users, utilizes a native large language model (LLM) capable of translating a combination of 16 different languages. The underlying model places a strong emphasis on maintaining accuracy, fluency, and contextual relevance to deliver the best possible translations.
In a recent announcement, Roblox’s Chief Technology Officer Daniel Sturman stated, “Using AI to automate real-time translations in text chat removes language barriers and brings more people together, no matter where they live in the world.” The translation tool currently accommodates English, Simplified and Traditional Chinese, French, German, Indonesian, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai, Turkish, Polish, and Vietnamese.
The translation feature activates automatically within chat boxes. When users communicate in different languages, the AI translates their messages in real time to the user’s default language. A translation icon appears in front of each sentence, allowing users to revert to the original language by clicking it.
While developing this tool, Roblox encountered a unique challenge: translating between one language and another was straightforward, but managing combinations of the 16 languages proved complex. Instead of creating 256 separate models (16 x 16), the company opted for an innovative solution. Sturman elaborated that they employed a unified, transformer-based architecture for the LLM, streamlining control over translation through the source sentence and target language. “This is like having multiple translation apps, each designed for a group of similar languages, all accessible via a single interface,” he explained.
The system introduces several distinctive features. It prioritizes accuracy, fluency, and contextual understanding while also translating sentences that combine various languages. To enhance the model’s performance, human evaluators contributed by teaching it contemporary slang and trending phrases in each of the 16 supported languages, a process that will be continually updated.
Furthermore, Roblox implemented a ‘back translation’ technique for less frequent translation pairs, like French to Thai. This method involves initially translating content and then translating it back to the original language to compare the two versions for discrepancies. The resulting labeled data is then used to enrich the model’s training dataset.
Looking ahead, Sturman indicated that Roblox is exploring the possibility of automatic voice chat translations for its users.
