Roblox Voice Safety Classifier
Open model flagging harmful speech in live voice chat.
Made by RobloxAI content classifiers
About Roblox Voice Safety Classifier
Roblox made a voice-safety classification model for voice-toxicity detection and classification. The latest voice-safety-classifier models are published under Roblox Corporation on Hugging Face. The model outputs have been calibrated for the Roblox voice chat environment.
The model supports eight languages: English, Spanish, German, French, Portuguese, Italian, Korean, and Japanese. It classifies input audio into six toxicity classes in a multilabel fashion.
Drafted from github.com · not yet confirmed by Roblox
How it works
This is a model that listens to short clips of voice chat and scores them for profanity, bullying, dating and sexual content, racism and other categories, released by Roblox with weights and code. It works on audio directly rather than transcribing first, which keeps latency low enough for live voice.
You host it yourself and feed it audio segments; it returns category probabilities you threshold. Roblox trained it on its own game voice data, so it's strongest on the kind of speech you hear in games with a young audience. Expect to fine tune for other domains.
Reviews
Be the first to review it
AnonymousKnow a team weighing Roblox Voice Safety Classifier? Send them this page.