A TinyML Approach to Real-Time Snoring Detection in Resource-Constrained Wearables Devices †

Abstract

This study proposes a health monitoring system for snoring detection utilizing Tiny Machine Learning (TinyML) models, specifically designed for resource-constrained wearable Internet of Things (IoT) devices. This research addresses significant constraints associated with running machine learning models on IoT devices, such as latency, limited memory, and low computational resources. These parameters are essential for real-time monitoring in healthcare applications, where prompt response is critical. The research focuses on developing a TinyML model capable of identifying specific audio patterns related to snoring during sleep. Experimental evaluations conducted in real-world sleep environments with the TinyML model deployed on resource-constrained wearable IoT devices. The evaluation results show that the proposed model achieves high accuracy while utilizing minimal computational resources and without introducing latency issues. Through the integration of audio (Syntiant) and advanced audio preprocessing techniques, the proposed system improves the efficiency of the TinyML model on wearable devices. The quantized TinyML model achieved an accuracy of 95.85% with a low latency of 48 ms, utilizing only 17.0K of RAM and 34.07K of flash memory for real-time snoring classification. This study highlights the benefits of practical deployment of the TinyML model for snoring detection on resource-constrained wearable IoT devices, demonstrating that such models can operate effectively within the constraints of current wearable technology.