AI-powered Emergency Keyword Detection for Autonomous Vehicles

Communication systems implemented within vehicles primarily focus on infotainment applications and often lack concrete specifications designed to ensure the safety of both drivers and passengers during emergencies. In this work, we propose and implement an in-vehicle emergency threat key-words detection system using machine learning and audio signal processing to classify potential emergency keywords based on audio waveforms while reducing noise and unprivileged user voice commands from creating a false positive. This is done by taking speech commands from users via a microphone array within the vehicle that is then preprocessed, and features are extracted from Mel spectrogram images. These mel spectrogram images are classified using a convolutional neural network (CNN) that has been previously trained to classify emergency speech keywords. Experimental results reveal a validation accuracy of about 90% is achieved in accurately detecting and classifying the threat keywords. The proposed system can be used in an emergency protocol within the autonomous vehicle by pulling over safely to prevent harm or call first responders.