Audio Driven Detection of Hate Speech in Telugu: Toward Ethical and Secure CPS

The rapid integration of social media platforms into cyber-physical systems has introduced new challenges in ensuring human-centric reliability and safety. This is mainly due to the widespread dissemination of hate speech and the inability of online systems to effectively moderate offensive content. While significant advances have been made toward hate speech detection in high-resource languages such as English, low-resource languages such as Telugu do not have the annotated datasets and tools to properly detect it. This project addresses this gap by creating a complete annotated multimodal hate speech dataset in the Telugu language, consisting of 2 h of audio-text pairs from YouTube. The dataset enables the exploration of hate speech detection in individual modalities, speech and text, as well as in a combined multimodal setting. The work presented in this paper is focused on the detection of hate speech based on audio data with text-based analysis incorporated as an ablation study to better understand the modality-specific contributions. Our classification results demonstrate that the combination of OpenSMILE acoustic features and an SVM classifier yields the highest performance in speech classification, achieving an F1 score of 0.89. In contrast, the best-performing text model, using LaBSE embeddings, attained an F1 score of 0.88.