Smart Home Security Through Real-Time Audio Event Detection Systems Using Convolutional Neural Network (CNN)

Smart homes increasingly demand intelligent security solutions that go beyond visual surveillance, yet Audio Event Detection (AED) remains underutilized. This paper presents a real-time AED system leveraging a Convolutional Neural Network (CNN) optimized for low-resource hardware. We first introduce TUTRARE, a novel security-focused audio dataset incorporating silence removal, volume normalization, and ambient noise variations. Our methodology applies Melspectrogram feature extraction (64 bands, $64 \times 64$ grid), data augmentation (noise addition, pitch shifts), and a compact CNN (two $3 \times 3$ convolutional layers with $32 \rightarrow 64$ filters, ReLU activations, max-pooling, and 30 % dropout). Training was conducted over 30 epochs using the Adam optimizer (learning rate $=0.001$) on a $70 / 20 / 10$ train/validation/test split. We evaluate performance using accuracy, precision, recall, F1score, ROC-AUC, and PR-AUC. The proposed model achieved 96.3 % accuracy and AUC $>0.99$, demonstrating robust detection of critical events such as glass breaking, gunshots, and sirens. Real-time deployment on Windows 10 (x64 CPU, 8 GB RAM) with an omnidirectional microphone confirmed end-to-end latency below 200 ms. These results indicate that audiobased security can effectively complement camera systems, addressing privacy blind spots and improving threat response in smart homes.