Advances in artificial intelligence have enabled synthetic voice generation systems that produce highly realistic human-like speech. While these innovations benefit multimedia production, accessibility, and creative applications, they introduce security vulnerabilities through voice spoofing, identity impersonation, and disinformation campaigns. This research presents a realtime detection framework to distinguish AI-generated speech from authentic human voices using machine learning implemented in MATLAB and Simulink. The system employs Mel-frequency cepstral coefficients combined with spectral features extracted from audio streams, processed through a Random Forest classifier to provide offline, low-latency detection. Experimental validation across diverse datasets demonstrates 92.3% detection accuracy with inference times below 200 milliseconds, establishing practical viability for securitycritical applications and fraud prevention.