SquawkBuster: Detecting Toxicity in Video-Game Voice Chat

The project aims to enhance toxic behavior detection in online multiplayer games through voice chat. While text-based detection models have advanced, voice chat presents unique challenges due to its complexity and resource demands; however, it offers valuable cues like tone and intonation. As such, the project seeks to develop multi-modal models combining voice and text data, optimized for gaming environments with limited resources. Leveraging Ubisoft’s expertise, the initiative aims to improve player safety by identifying toxic behavior more effectively. The project has two main objectives: (1) developing and evaluating models that analyze both voice and text interactions, and (2) optimizing these models for low-resource environments. The research will construct a dataset using open-source and Ubisoft data, comparing models like MuTox and ToxBuster in multi-modal settings. Moreover, the inclusion of optimization techniques such as model distillation and wake-word triggering will reduce computational load while maintaining accuracy. Additionally, spoken term discovery will identify game-specific vocabulary to improve classification and alert potential toxic behavior. The expected deliverables include insights and tools for advanced voice-based toxicity detection systems, enhancing player safety, improving gaming quality, fostering safer communities, and reducing moderator burden. Taken together, this work will advance industry standards in digital safety in gaming.

Faculty Supervisor:

Ewan Dunbar

Student:

Partner:

Ubisoft Toronto

Discipline:

Sociology

Sector:

Information and cultural industries

University:

University of Toronto

Program: