Classification of emerging sounds during the sleep

Sounds like snoring, coughing, sneezing, whistling, which have different acoustic properties, can emerge involuntarily during the sleep. These sounds may affect negatively the sleep quality of the other people in the same environment, just as it may affect directly the sleep quality of the person. To increase the sleep quality of the person, these sounds should be recorded and these records should be evaluated by a sleep expert. The appropriate treatment method should be determined according to the evaluation result. However, since the duration of a normal sleep is in the range of six hours to eight hours, it's a time consuming process for the sleep expert to listening all records. Therefore, it has been determined that a timesaver algorithm which analyzes automatically the sounds recorded during the sleep, is needed to help diagnose of the sleep expert .In the previous literature review, it has been observed that the number of studies on this topic is very few. Majority of existing studies detects only the snoring sounds, but these works do not involve the algorithms which help the sleep expert to diagnose. In this study, differently from the literature, it has been aimed that developing a computer aided diagnosing algorithm which will classify the sounds emerging during the sleep automatically with high accuracy by analyzing the all records in a fast and effective way to help the sleep expert to diagnose. In accordance with this purpose, at the first stage, sleep sounds were researched and different types of sound recordings obtained from open access sites thereby internet browsing. All records have been listened one by one and have been grouped. In the second stage of the study, the methods mentioned in the literature have been validated. Within this period, energy, variance, zero cross rate, autocorrelation function, Fourier transform methods have been applied on the all records and the features which will provide general information about records, have been obtained. In the third stage of the study, an adaptive segmentation algorithm has been developed and records has been segmented automatically as every segment has one kind of sound. The acoustic features belonging to analyzed sound have been obtained by applying analytical methods to each segment. In addition, the visual and mathematical features have been obtained in frequency, time and amplitude domains by applying "Continuous Wavelet Transform" for the every segment. At the last stage, support vector machine used as a classifier and it has been trained with the obtained features. The success of the algorithm has been tested by using different records. 390 and 449 segments were used for training and testing respectively. As a result of the study, developed algorithm has classified six different parameters which are exhalation, simple snoring, high frequency duplex snoring, low frequency duplex snoring, triplex snoring and coughing, with 96.44% accuracy rate.