Automatic snoring sounds detection from sleep sounds based on deep learning

摘要

Snoring is a typical characteristic of obstructive sleep apnea hypopnea syndrome (OSAHS) and can be used for its diagnosis. The purpose of this paper is to develop an automatic snoring detection algorithm for classifying snore and non-snore sound segments, which have been segmented from a whole-night sleep sound signal using a spectral entropy method, based on convolutional neural network (CNN) descriptors extracted from audio maps. For each sound segment, the time-domain waveform, spectrum, spectrogram, Mel-spectrogram and CQT-spectrogram are calculated. Two classifiers are applied to classify sound segments into either snore or non-snore classes. The first classifier is referred to as CNNs-DNNs and combines CNNs and deep neural networks (DNNs), and the second classifier is referred to as CNNs-LSTMs-DNNs and consists of CNNs, Long and Short memory networks (LSTMs) and DNNs. The results show that the Mel-spectrogram can better reflect the differences between snore and non-snore sound segments for the five maps extracted in this study. Furthermore, the deep spectrum features extracted from CNNs-LSTMs-DNNs using Mel-spectrogram are well suited to this task. The results indicate that the method developed in this study could be used for a portable sleep monitoring device.