Audio-Based Hate Speech Detection Using Speech Recognition and Transformer Techniques

View at publisher Summarise with VIRAT
Authors: Manar Ali Alsehaimi, Abdullah Alsaeedi, Wael M. S. Yafooz
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Hate speech poses significant challenges in online communication by spreading harmful content that discriminates against individuals or groups. This study proposes a novel framework for detecting hate speech from spoken Saudi dialect by combining Automatic Speech Recognition (ASR) and transformer-based deep learning models. We developed a custom Saudi hate speech audio dataset of 1,700 samples evenly balanced between offensive and non-offensive speech collected from native speakers. Audio samples were classified with five Arabic transformer models: AraBERT-v2, MARBERT, QARiB, SaudiBERT, and CAMeLBERT-DA. We evaluated both ASRgenerated transcripts prone to recognition errors and manually verified gold transcripts to measure model robustness against ASR quality. Experimental results show that despite a moderate ASR Word Error Rate (WER) of 39.38%, transformer models achieve strong classification performance, with MARBERT and SaudiBERT attaining over 96% F1-score on ASR transcripts and up to 100% on gold transcripts. The ensemble of models further improves classification stability.

Fields of Study & Subjects

Hate Speech and Cyberbullying Detection Speech Recognition and Synthesis Sentiment Analysis and Opinion Mining

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
1335–1341
Publication date
21 January 2026
Language
EN