AcBreath: In-Hand Free Acoustic Breath Signal Reconstruction Using Variational Autoencoder

Acoustic signals from smartphones have become a fascinating research topic for human sensing. However, acoustic signals face a limitation when detecting chest signal vibrations (i.e., respiration, heart rate, and blood pressure) during in-hand activities (between smartphone and hand manipulation). The chest signal variation mixes with the signal variation induced by the signal reflection from the hand and other body parts, which may result in signal loss. In this article, we propose AcBreath, a solution to address the signal loss of respiratory vibrations. We exploit the “dual-mode acoustic signals” detection mode to capture the dynamic signal reflection, the environmental signal, and the ambient noise signal. We establish a resilient and optimized signal processing that involves the proposed frequency short-time energy (FSTE) to infer low-energy signal events, such as weak respiratory reflections drowned by ambient noise. Thus, the inferred signal is processed and reconstructed using the proposed self-supervised variational autoencoder (VAE) network, called AcBreathNet (AcBNet), to obtain an accurate respiratory rate. Comprehensive experiments demonstrate the effectiveness of AcBreath by achieving an overall error rate of 0.275 bpm on the dataset.