Enhancing Ride-Hailing Safety through Real-Time Speech-Based Violence Detection

The rise of ride-hailing services has brought growing safety concerns, especially incidents of verbal harassment during trips. While prior research has focused mainly on visual-based violence detection, this study addresses the underexplored area of real-time speech-based harassment detection. We present a multimodal safety framework that integrates OpenAI's Whisper for speech transcription with a fine-tuned DistilBERT model for toxicity classification, trained on the Jigsaw Toxic Comment Classification dataset. Our system achieves an impressive 93.8% accuracy, surpassing current state-of-the-art methods in toxic speech detection. While real-time capability is demonstrated through system design and latency evaluation, large-scale field trials remain future work. Designed for real-time processing, the framework enables proactive safety monitoring, making it ideal for ride-hailing and similar dynamic urban environments. This work contributes to the field by effectively combining automatic speech recognition and natural language processing for real-world safety applications. By bridging the gap between static datasets and live environments, our approach offers a practical, scalable, and impactful solution for enhancing passenger safety through real-time verbal abuse detection.