USING MICROPHONE AND ML TO DETECT THE PRESENCE OF HUMANS IN SPACE

Abstract

This project explores the development and performance of a voice recognition system implemented on an ESP32 microcontroller, utilizing a 1D Convolutional Neural Network (CNN) architecture. The system’s objective is to detect human presence by recognizing individual vocal characteristics through real-time audio input. The research extends into quantization techniques, employing the EON compiler to optimize the CNN model for efficient execution on the constrained hardware, reducing memory and flash usage while maintaining accuracy. The system was evaluated on a dataset split into training and testing subsets, achieving a remarkable accuracy of 90.72% on the testing set, surpassing the initial accuracy target of 80% set during the project’s inception. The integration of the MAX9814 microphone with the ESP32’s Direct Memory Access (DMA) and built-in I2S protocols enabled high-fidelity audio recording without delays. This project not only confirms the feasibility of deploying machine learning models on low-resource microcon- trollers but also provides a foundation for future enhancements in biometric-based security and personal identification systems.