Roblox Voice Safety Classifier

weights

open(Apache-2.0 safetensors on HF Roblox/voice-safety-classifier-v3, with no access gate)

data

described(Roblox's internal voice chat, machine- and human-labeled, described in the v3 card and announcement but not released)

code

partial(an inference script and requirements files only, no training code)

license

Apache-2.0(OSI, the v3 card and its LICENSE.md)

prior-release

CC-BY-SA-3.0(v1 (Roblox/voice-safety-classifier) and v2 (Roblox/voice-safety-classifier-v2) stay downloadable under CC-BY-SA-3.0. v3 is the current release, so its license governs)

Version 3 is published under Apache 2.0 with no access gate and an inference script, while its training data, which Roblox describes, and its training code are not released. The earlier v1 and v2 checkpoints carry a Creative Commons share-alike license, but v3 is the current release, so its terms decide.

• https://about.roblox.com/newsroom/2026/06/upgrading-voice-safety-classifier-22-languages-sharper-detection-capabilities recorded 2026-10-19

  Roblox's v3 announcement of June 17, 2026 says "we used 250,000 hours of machine-labeled multilingual data and 29,000 hours of human-labeled multilingual data", and links no training data.

• https://huggingface.co/api/models/Roblox/voice-safety-classifier recorded 2026-10-19

  The API record for the v1 repository carries the tag "license:cc-by-sa-3.0" and is not gated.

• https://huggingface.co/api/models/Roblox/voice-safety-classifier-v2 recorded 2026-10-19

  The API record for the v2 repository carries the tag "license:cc-by-sa-3.0" and is not gated.

• https://huggingface.co/api/models/Roblox/voice-safety-classifier-v3 recorded 2026-10-19

  The Hugging Face API record reports "gated":false and the tag "license:apache-2.0", and lists "model.safetensors", "inference.py", "requirements.txt" and "LICENSE.md" among its files, with no training script.

• https://huggingface.co/Roblox/voice-safety-classifier-v3/raw/main/LICENSE.md recorded 2026-10-19

  The LICENSE.md in the v3 repository is the Apache License, Version 2.0 text, "TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION".

• https://huggingface.co/Roblox/voice-safety-classifier-v3/raw/main/README.md recorded 2026-10-19

  The v3 card front matter declares "license: apache-2.0" and its License section reads "Apache License 2.0", pointing to LICENSE.md. It says "All the model training has been conducted with Roblox internal voice chat datasets, using both machine and human-labeled data, with 300k hours of training data in total", and its usage section runs "python inference.py"; it mentions no training script.

Adoption is measured by Hugging Face downloads of all three releases, v3 and the earlier v1 and v2. Downloads of the older weights from the GitHub repository's release pages and Roblox's own use on its voice chat are not counted, so the figure is a lower bound on use.

• https://huggingface.co/api/models/Roblox/voice-safety-classifier recorded 2026-10-19

  93 downloads in the trailing 30 days for Roblox/voice-safety-classifier

• https://huggingface.co/api/models/Roblox/voice-safety-classifier-v2 recorded 2026-10-19

  178 downloads in the trailing 30 days for Roblox/voice-safety-classifier-v2

• https://huggingface.co/api/models/Roblox/voice-safety-classifier-v3 recorded 2026-10-19

  190 downloads in the trailing 30 days for Roblox/voice-safety-classifier-v3

It moderates speech directly, across 30 languages, where Llama Guard reads text. Its eight fixed categories cover fewer harms than Llama Guard's fourteen, and on Roblox's own tables it still misses a large share of violations in most languages, so it sits a step below Llama Guard.

• https://about.roblox.com/newsroom/2026/06/upgrading-voice-safety-classifier-22-languages-sharper-detection-capabilities recorded 2026-10-19

  The announcement says "V3 of the open-source model achieves 61% recall weighted by Roblox voice chat language distribution at 1% false positive rate", and adds 22 languages and two violation categories over v2.

• https://huggingface.co/Roblox/voice-safety-classifier-v3/raw/main/README.md recorded 2026-10-19

  The v3 card lists eight heads from "ABUSE_TYPE_PRIVACY_ASKING_FOR_PII" to "ABUSE_TYPE_DISRUPTIVE_AUDIO", says "The model supports 30 languages" with "auxiliary heads for language detection", sets "The maximum supported audio duration is 30s", and gives recall at a 1% false-positive rate from 29.5% for Czech to 75.2% for Polish.