Snoring-related Signal Automatic Detection And Classification

For acoustic based methods on diagnosing the snore source and obstruction site in the upper airway of Obstructive Sleep Apnea/Hypopnea Syndrome (OSAHS) patients, it is essential to implementing automatically detecting and classifying different snore related signals (SRS) data from the microphone audio recording. Specifically, for doctors and researchers on this area to carry on establishing SRS database and deep study, a software framework for automatic detection and classification of different SRS data could be significant to enhance the efficiency of work. Interests in modern signal processing, have expanded from classical filtering and spectrum analysis to new paradigms like machine learning and intelligent information processing. Therefore it is possible to provide sufficient theoretical fundamentals to implementing automatically detection and classification of different SRS data based on the combination of traditional signal processing techniques with machine learning. Relevant algorithms and methods in this thesis could also be use into some other applications like microphone based smart home systems, intelligent language recognition systems, instruments recognition systems, etc.We are talking in this thesis on how to establish a whole framework for automatically detecting and classifying different SRS data from microphone audio recording. The main work of this thesis is to propose a whole practical method to detect and classify different SRS data automatically, an audio recording of snoring has been used to demonstrate the efficiency of methods we proposed. We adopted the concept of missing alarming rate and false alarming rate into the detection of SRS events, which has been proved to achieve a good performance in experiments. In addition, we extracted and compared varies of acoustic features from the SRS data. Then after extraction of features we utilize an algorithm to select the most useful features, which could be helpful for reducing the dimensions of features for machine learning. Subsequently, we selected the best trained classifier by validation method. Finally, we embedded the trained classifier into our detection processing framework, leading to a whole software framework for automatic detection and classification of different SRS data. The methods we proposed in this thesis are demonstrated to be available both in the experiments and subjective testing. In addition, we utilized the software framework to automatically detect and classify inspiration related snoring signals and non-inspiration related snoring signals. The efficiency of our method was demonstrated by both the theoretical analysis and the simulation experiments.