Snore Sound Recognition via an Explainable Capsule Network

Snoring can be caused by an upper airway reaction while sleeping. Classifying the excitation locations of snore sounds accurately provides assistance for treating snoring. In this research, we propose a convolutional neural network combined with a Capsule Network (CapsNet) to solve this problem. The models were trained and tested on the Munich-Passau Snore Sound Corpus (MPSSC), a relatively small and imbalanced dataset that contains four classes. As a result, the proposed method achieved an Unweighted Average Recall (UAR) of 58.5 %. Furthermore, we explained the working principle of the CapsNet through visualization, which could be helpful for understanding the generation of the results.

I. Introduction

Snoring is a prevalent disorder that affects 20-40 % of the general population [1]. There are various causes of snoring, some of which are due to vibrations in the soft tissue structures of the upper airway caused by inspiratory airflow. One type of snoring, obstructive sleep apnea, is featured as a sleep disorder in this study. Most importantly, this sleep disorder could lead to hypercoagulation and cardiovascular disease [2]. Computer audition has been shown that audio data can be efficient for symptom analysis in the medical field [3]. For this purpose, this research uses Capsule Network (CapsNet). Dynamic routing algorithms, implemented within CapsNet, have found applications in various domains, including the healthcare sector. It can maximize the matching between capsule vectors and encode visually explainable numerical features. For this reason, CapsNet is used to implement an explainable model.