Automatic detection of snoring events using Gaussian mixture models

in this work, an automatic snore detection system of acoustic snoring signals has been designed. its purpose is to assist an alternative non-invasive method for diagnosing obstructive sleep apnea (OSA) based on acoustic signal processing. the detector is based on Gaussian mixture models that were trained and validated on full night acoustic signals that were recorded from a sleep laboratory, along with polysomnographic tests taken from patients with widely distributed severity of OSA. the snore detection system includes steps from noise reduction through event detection and all the way to snore identification. in order to analyze the performance of our proposed detector, a total of more than 80,000 acoustic episodes from 33 different OSA patients were manually segmented into snore and non-snore episodes; among the non-snore episodes we can find a variety of sleep related noises such as blanket and pillow murmurs, moaning, groaning, coughing, and talking. the validation dataset was recorded using two different audio recorders to ensure the robustness of the detector. the events' total identification rate was 97.12% with 96.02% positive detection of snore as snore (sensitivity) and 97.90% detection of noise as noise (specificity).