Application of Artificial Neural Networks for Audio Classification on an Embedded Platform

The paper investigates the application of state-of-the-art artificial neural networks (ANNs) for audio classification tasks on embedded systems, specifically the NVIDIA Jetson Nano. The authors focus on evaluating the performance of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers in classifying unmanned aerial vehicle (UAV) noise. Key challenges in this domain, such as high levels of noise and variability in input data, limited labeled audio datasets, and the need to adapt ANNs to resource-constrained embedded systems, are discussed. The study delves into the architectural characteristics of these ANN models, highlighting the strengths of CNNs for spectrograms, RNNs with LSTM and GRU for sequential data, and Transformers for their ability to handle long-range dependencies. A methodology for implementing these models on the Jetson Nano platform is proposed, considering its hardware constraints. Experimental results demonstrate the effectiveness of different architectures for UAV noise classification, with CNNs excelling in spectrogram analysis, while RNNs and Transformers proving more suitable for raw audio or sequential feature processing. The authors outline directions for future research, including the development of optimised training methods for small datasets and the adaptation of advanced noise classification approaches to resource-constrained platforms. Recommendations are provided for selecting the appropriate architecture based on task-specific requirements and constraints.