Environmental sound recognition on embedded devices using deep learning: a review

Sound recognition has a wide range of applications beyond speech and music, including environmental monitoring, sound source classification, mechanical fault diagnosis, audio fingerprinting, and event detection. These applications often require real-time data processing, making them well-suited for embedded systems. However, embedded devices face significant challenges due to limited computational power, memory, and low power consumption. Despite these constraints, achieving high performance in environmental sound recognition typically requires complex algorithms. Deep Learning models have demonstrated high accuracy on existing datasets, making them a popular choice for such tasks. However, these models are resource-intensive, posing challenges for real-time edge applications. This paper presents a comprehensive review of integrating Deep Learning models into embedded systems, examining their state-of-the-art applications, key components, and steps involved. It also explores strategies to optimise performance in resource-constrained environments through a comparison of various implementation approaches such as knowledge distillation, pruning, and quantization, with studies achieving a reduction in complexity of up to 97% compared to the unoptimized model. Overall, we conclude that in spite of the availability of lightweight deep learning models, input features, and compression techniques, their integration into low-resource devices, such as microcontrollers, remains limited. Furthermore, more complex tasks, such as general sound classification, especially with expanded frequency bands and real-time operation have yet to be effectively implemented on these devices. These findings highlight the need for a standardised research framework to evaluate these technologies applied to resource-constrained devices, and for further development to realise the wide range of potential applications.