Back office ops · Production

Roblox builds a unified multilingual translation model for real-time cross-language chat

The problem

Roblox's global platform of more than 70 million daily active users communicating in their native languages across more than 15 million active experiences could not be served by manual translation, and existing AI translation only covered static content rather than real-time chat interactions.

Workflow diagram · grounded in source
1
User sends in-experience chat
trigger
“In any experience that has enabled our in-experience text chat service, people from different countries can now be understood by people who don't speak their language”
2
Unified LLM translates message
ai_action
“we built a unified, transformer-based translation LLM to handle all language pairs in a single model. This is like having multiple translation apps, each specializing in a group of similar languages, all available with a single interface…”
3
Trust and safety check
validation
“This back end is where we apply additional chat translation logic and integrate the system with our usual trust and safety systems. This ensures translated text gets the same level of scrutiny as other text, in order to detect and block …”
4
Translated message delivered
output
“The chat window will automatically show Korean translated into English, or Turkish translated into German, and vice versa, so that each person sees the conversation in their own tongue. These translations are displayed in real time, with…”
5
Quality evaluation improves model
feedback_loop
“In order to assess quality, we built an ML model and trained it on human labeled error types and scores. We then fine-tuned a multilingual language model to predict word-level errors and types and calculate a score using our multidimensi…”
Reported outcome

Roblox deployed real-time AI chat translations across 16 languages with approximately 100ms latency, capable of handling more than 5,000 chats per second, and the model outperforms commercial translation APIs on Roblox content while driving stronger engagement and session quality.

Reported metrics
Translation latencyapproximately 100 milliseconds
Chats per second supportedmore than 5,000
translation quality vs commercial APIsoutperforms commercial translation APIs on Roblox content
Engagement and session qualitystronger engagement and session quality
Show all 7 reported metrics
translation latencyapproximately 100 milliseconds
chats per second supportedmore than 5,000
translation quality vs commercial APIsoutperforms commercial translation APIs on Roblox content
engagement and session qualitystronger engagement and session quality
daily active users on platformmore than 70 million
active experiences on platformmore than 15 million
distilled model sizefewer than 650 million parameters
Reported stack
LLMML model
Source
https://corp.roblox.com/newsroom/2024/02/breaking-down-language-barriers-with-a-multilingual-translation-model
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Roblox deployed real-time AI chat translations across 16 languages with approximately 100ms latency, capable of handling more than 5,000 chats per second, and the model outperforms commercial translation APIs on Roblo…

What tools did this team use?

LLM, ML model.

What results were reported?

Translation latency: approximately 100 milliseconds; Chats per second supported: more than 5,000; translation quality vs commercial APIs: outperforms commercial translation APIs on Roblox content; Engagement and session quality: stronger engagement and session quality (source-reported, not independently verified).

How is this back office ops AI workflow structured?

User sends in-experience chat → Unified LLM translates message → Trust and safety check → Translated message delivered → Quality evaluation improves model.

Related back office ops cases