The Study Of Cough Signal Detection Based On Gaussian Mixture Model

Cough is one of the most common symptoms of many respiratory diseases. When frequent, severe and persistent cough appears,detecting the intensity and frequency of cough through analysis of cough sounds can provide much valuable clinical messages, which are useful for diagnosing the disease and assessing the effectiveness of treatment quantitatively.?However,the patient's perception of the symptom may be incomplete and inaccurate, which makes the diagnosis and treatment of diseases difficult. The subjectivity and lack of robustness of the assessment have led to increased interest in developing automated ambulatory cough monitors as an objective method of measuring the frequency and intensity of cough in patients suffering from chronic cough.? This requires an algorithm capable of detecting the majority of cough sounds present in a given recording while rejecting other sounds that have similar characteristics to cough sounds, which is the key and foundation of future research.This paper points out the differences and relationship between cough sound detection and speech recognition. Based on a comprehensive investigation of the characteristic of cough sound and technologies on speech recognition, a detection system of cough signals based on Gaussian Mixture Model is built on MATLAB platform and the application of wavelet analysis theory in this detection system is researched in detail.?The concrete content is as follows:①At the respect of pretreatment, this paper especially studies a noisy cough sound signal denoising algorithm based on wavelet transform, through plentiful experimental work and elaborate analysis of the key technical problems such as the selection of the wavelet and decomposition level,threshold processing method, the suitable wavelet function is selected and wavelet threshold denoising method is used to suppress noise; then we use double threshold detection method based on short-time energy and short-time zero crossing ratio to implement the endpoint detection of cough. The experiment shows the robustness of cough sound detection system is improved in noisy environment and the amount of computation is also reduced.②This thesis analyzes three effective characteristic features uesd for analysis and extraction of audio signals: Linear Prediction Coefficients,Linear Prediction Cepstrum Coefficients and Mel Frequency Cepstrum Coefficients. Combined with wavelet packet transform and Mel Frequency Cepstrum Coefficients, we extract the WPT-MFCC parameters. The experimental results show that the WPT-MFCC has better recognition accuracy in noisy circumstances compared with other three parameters.③On the basis of studying the algorithm of expectation maximization for model parameter estimation,establishing method of classification model base and judgment standard of recognition results, we create corresponding Gaussian Mixture Models for the cough sound, speech voice, laughter and throat clearing sound in the recordings respectively. Then simulation experiments are used to analyze the effect of denoising and the performance of system in different model orders, methods of feature extraction, we obtain a better detection system which can reach higher recognition rate and lower error rate.This paper makes a basic and preparatory study for exploring further practical detection and analysis system of cough sounds by investigating and testing cough sound detection system based on GMM.