Towards automatic pitch detection in snoring signals

Natural snores are nocturnal breath sounds that can be captured and studied with signal processing techniques. The fundamental frequency of snores has been used for snore classification, according to the site of obstruction in the upper airway, and can be useful for the early diagnosis of some diseases such as Multiple System Atrophy. The relationship between fundamental frequency of snores and the Obstructive Sleep Apnoea Syndrome has not been studied. The pitch is the time domain counterpart of the fundamental frequency. It can be defined for snoring signals as it is for speech vowels. Many automatic pitch detection methods have been developed for speech. In this work one of the oldest pitch detection algorithms, namely the autocorrelation detector with centre clipping, is adapted to snores. A manual pitch estimation method is proposed for evaluating the performance of the automatic detector. Both methods are applied to some snores from simple snorers and from patients with OSAS. The automatic algorithm, which follows slow pitch variations during a snore, produces a smoothed version of the manual pitch estimation. The automatic algorithm also allows the detection of pitch absence. Three parameters-pitch mean value, pitch standard deviation and pitch density-are defined for the analysis of pitch evolution into a snore. These parameters show similar values for manual and automatic algorithms. Some differences in pitch density are found between simple snorers and OSAS patients. Those differences need to be confirmed over a more exhaustive database