Emotion and violence detection from environmental sounds based on next generation machine learning methods

This thesis presents a comprehensive framework for the development of lightweight, interpretable, and high-performance machine learning models for various audio-based recognition tasks, such as environmental audio classification, speaker counting, intensity detection, and ensemble emotion recognition. Seven new models are proposed, each incorporating handcrafted, lightweight, pattern recognition or chaos-inspired feature extraction techniques, optimized classification strategies, and multi-level signal transformations such as DWT and TQWT. Each model is trained and evaluated on novel, balanced and publicly available datasets specifically created for their respective tasks. All models run without the need for GPU acceleration, demonstrating that high accuracy audio classification is possible with low computational resources. Overall, this thesis presents a powerful alternative to deep learning by providing explainable, efficient and scalable models suitable for use in real-time, embedded or low-resource environments. This thesis makes a significant contribution to the field of emotion and violence detection with next generation machine learning models. The proposed solutions pave the way for advanced applications in public safety, forensics, intelligent surveillance, health monitoring and emotional computing. The proposed solutions pave the way for advanced applications in public safety, forensics, intelligent surveillance, health monitoring and emotional computing.