Convolutional Neural Networks for audio classification on ultra low power IoT devices

Sound classification usually requires heavy resources in terms of computation, memory, and energy to achieve good accuracy. However, it is possible to enable a more efficient and accurate audio recognition on pervasive IoT platforms through specific optimizations. In this paper, a solution based on convolutional neural networks is proposed for audio classification on resource-constrained wireless edge devices. Furthermore, different pre-processing techniques have been tested to evaluate the classification accuracy with respect to computational, memory, and energy footprint.