Unsupervised Snore Detection from Respiratory Sound Signals

Acoustic monitoring and diagnosis of Obstructive Sleep Apnea (OSA) has attracted increasing attentions recently due to its advantages over traditional Polysomnography (PSG) based diagnosis. The performances of several existing approaches are often limited by several aspects such as large amount of training data and and high computational cost. In this paper, we proposed a novel snore detection method utilizing the Vertical Box (V-Box) algorithm to detect snore candidate, and then we calculate the spectral feature Mel-Frequency Cepstral Coefficients (MFCCs) of each candidate. Finally, K-Harmonic Means (KHM) clustering algorithm are used to achieve the general classification decision (snore or non-snore). Experimental results demonstrate that the overall accuracy of our proposed method is 95.6% with Positive Predictive Value (PPV) 96.0%, which is superior to the existing approaches.