ToxiAlert-Bench: Audio Dataset for Toxic Speech Detection with Paralinguistic Cues
A recent study published on arXiv presents ToxiAlert-Bench, a comprehensive audio dataset aimed at detecting toxic speech, which includes paralinguistic elements such as emotion, intonation, and speech rate. This dataset features more than 30,000 audio recordings categorized into seven primary toxic categories and twenty detailed toxic labels. Notably, it differentiates between sources of toxicity from textual and paralinguistic aspects. Additionally, the researchers introduce a dual-head neural network designed with a multi-stage training approach specifically for identifying toxic speech. This research tackles the shortcomings of current text-focused datasets and models that overlook paralinguistic factors.
Key facts
- ToxiAlert-Bench is a large-scale audio dataset for toxic speech detection.
- Dataset includes over 30,000 audio clips.
- Annotated with seven major toxic categories and twenty fine-grained toxic labels.
- Distinguishes toxicity sources: textual content vs. paralinguistic origins.
- Proposes a dual-head neural network with multi-stage training strategy.
- Addresses neglect of paralinguistic cues in existing approaches.
- Published on arXiv with identifier 2605.15984.
- Focuses on toxic speech detection in online communication.
Entities
Institutions
- arXiv