SynHate: Detecting Hate Speech in Synthetic Deepfake Audio

View on arXiv ← Back to list

Authors: Rishabh Ranjan, Kishan Pipariya, Mayank Vatsa, Richa Singh

Published: 2025-06-07 11:46:39+00:00

AI Summary

SynHate is the first multilingual dataset for detecting hate speech in synthetic audio, encompassing 37 languages and a novel four-class scheme (Real-normal, Real-hate, Fake-normal, Fake-hate). It leverages pre-trained self-supervised models to evaluate hate speech detection performance, revealing variations across languages and highlighting the challenge of cross-dataset generalization.

Abstract

The rise of deepfake audio and hate speech, powered by advanced text-to-speech, threatens online safety. We present SynHate, the first multilingual dataset for detecting hate speech in synthetic audio, spanning 37 languages. SynHate uses a novel four-class scheme: Real-normal, Real-hate, Fake-normal, and Fake-hate. Built from MuTox and ADIMA datasets, it captures diverse hate speech patterns globally and in India. We evaluate five leading self-supervised models (Whisper-small/medium, XLS-R, AST, mHuBERT), finding notable performance differences by language, with Whisper-small performing best overall. Cross-dataset generalization remains a challenge. By releasing SynHate and baseline code, we aim to advance robust, culturally sensitive, and multilingual solutions against synthetic hate speech. The dataset is available at https://www.iab-rubric.org/resources.

Key findings

Whisper-small demonstrated the best overall performance across languages. Significant performance differences were observed across languages and between datasets, highlighting the challenge of cross-dataset generalization in multilingual hate speech detection. Cross-dataset accuracy rarely exceeded 52%.

Approach

The research uses five pre-trained self-supervised models (Whisper-small/medium, XLS-R, AST, mHuBERT) to evaluate the SynHate dataset. These models are fine-tuned on the dataset for a four-class hate speech detection task, and their performance is analyzed across different languages and datasets.

Datasets

SynHate (created from MuTox and ADIMA datasets)

Model(s)

Whisper-small, Whisper-medium, XLS-R, AST, mHuBERT

Author countries

India

← Previous