Neural Network-Based Gunshot, Chainsaw, and Forest Sound Detection on Embedded Monitoring Platforms

This paper presents a modular neural network (NN) classifier integrated into the SiRPA system for the recognition of acoustic patterns. The proposed design replaces hidden Markov models (HMMs), achieving superior accuracy, faster convergence, and efficient operation on low-power embedded devices. Each class (chainsaw, gunshot, forest) is handled by a binary multilayer perceptron with two hidden layers and tanh activations. The NN achieved validation accuracies of 97.9% for chainsaw and gunshot, and 93.2% for forest, surpassing HMM baselines by more than 8 percentage points. Convergence was also faster, with chainsaw and gunshot reaching mean squared error (MSE) below 10^-7 in fewer than 200 epochs. Deployment on Arduino Nano 33 BLE Sense, ESP-EYE, and Sony Spresense confirmed real-time inference with average power consumption in the tens to hundreds of milliwatts. These results establish modular NNs as accurate, efficient, and low-cost alternatives to HMMs for embedded acoustic monitoring.