LightSnore-Net: A Lightweight Neural Network for Snoring Detection and Mitigation in Smart Pillows

Snoring, a common sleep disorder, is often delayed in diagnosis due to the inconvenience of polysomnography (PSG). Current anti-snoring devices on the market often employ invasive designs that not only compromise wearer comfort, but can also disrupt normal sleep patterns. To address this issue, our study leverages recent advances in deep learning and Internet of Things (IoT) sensor technologies to propose a novel snoring detection network, LightSnore-Net. This network, based on a lightweight attention mechanism, can be deployed in the STM32F103C8T6 microcontroller of a smart pillow to facilitate non-invasive snore monitoring in a home environment. By analysing the Mel Frequency Cepstral Coefficients (MFCC) features of snore sounds extracted from audio modules and integrating a lightweight channel attention mechanism with a convolutional neural network, LightSnore-Net is able to accurately identify key features of snore sounds. On public datasets, this model achieves 99% detection accuracy, significantly outperforming traditional methods and recent deep learning models. This research provides an efficient and accurate solution, paving a new way for early diagnosis and monitoring of snoring.