Research On Detection And Analysis Algorithm Of Snoring Signal

Snoring is a common disease affecting 20%-40% of the general population. Snoring is not only a source of irritation to the patient and observer, but also a threat to the health of the patient. Obstructive sleep apnea syndrome is a kind of respiratory diseases, and it is usually associated with severe snoring. Obstructive sleep apnea may lead to excessive daytime somnolence and fatigue, as well as cardiovascular diseases. Polysomnography is considered to be the most reliable technique for the diagnosis and assessment of snoring. However, it requires numerous physiologic electrodes to be attached onto the patient for an overnight stay in a sleep laboratory. Due to its potential advantages, such as noninvasiveness, low cost, and simple to apply, acoustical analysis of respiratory sound signals has received considerable attention and research. The entire night recording has long time duration, and is often accompanied by non-snore audio events. Hence, developing an automatic snore detection method to analyze fullnight recordings in a timely and accurate manner would be advantageous. The analysis of snoring may also disclose additional clues relevant to clinical practice.In this paper, a novel unsupervised automatic snore detection algorithm is proposed. For the detection procedure, two main steps, segmentation and classification, are involved. Vertical box, which is a quick detection technique to solve the change-point problem, is adopted and modified to detecte and segmente the audio events in the fullnight recordings adaptively. As such, it will be coined as the adaptive vertical box algorithm. After segmentation, all the audio events will be represented as Mel-frequency cepstral coefficients and clustered by k-harmonic means algorithm in order to classify them into snore or no-snore classes. Experiment results demonstrate that this novel snore detection algorithm has an excellent performance.In this paper, a new snore analysis algorithm based on acoustic features is proposed. Snoring sound is characterized by a specific set of formants, by which this paper research the changes of snoring in the entire night. The snore events are initially extracted from the fullnight recording, and then they will be characterized as a set of formants and clustered by k-medoids algorithm. In order to achieve a better performance, k-medoids algorithm is modified in this paper. Experiment results show the changes of snoring in the whole night, and fullnight snoring sounds can be divided into three classes.