Recognition And Study Of Cough Sound Based On Hilberr Huang Transform

Cough is one of the most common symptoms of respiratory system disease. Its parameters--time of duration, type, intension and frequency provide significant information to clinical research. However, at present, cough just can be estimated based on patients’ subjective description resulting in failing to estimate cough objectively and quantificationally. With the development of computer and speech recognition, people is anxious for exploiting an automatic system for cough sound detection and recognition based on computer. This system can check out and recognize cough sound from audio stream, which contains amounts of background sounds and the sounds sharing similar properties. Thanks to this function, it is possible to objectively and quantificationally evaluate the features of cough, such as type, intension and frequency.Among the techniques of automatic detection and recognition of cough sound, the improvement of effect of cough sound recognition lies on the effective method of feature extraction and valid process of recognition. For the difference of occurrence mechanism between cough sound and speech, the former being typical non-stationary signal with evident characteristics of burst and turbulence, this paper attempts to adopt Hilbert Huang Transform(HHT) to extract some new feature parameters of cough sound, which is more suitable to process non-linear and non-stationary signal.The main content of this topic research is fusing the signal analysis methods of HHT to explore certain new feature parameters of cough sound. A total of560cough sounds were classified into two groups:280normal and280pathological, of which150were used for training respectively, and the rest for testing. Moreover,560non-cough sounds were classified into five groups:120throat clearing,120sigh voice,100shouting voice,120speech voice and100laughter, respectively. And then60of each were chosen for training, the rest for testing. This paper selected Hidden Markov Model(HMM) to implement the recognition of cough sound and simulated the experiment in MATLAB platform.First of all, this paper extracted Mel Cepstral Coefficients(MFCC) and modified Mel Cepstral Coefficients(MFCC1), then combined these two coefficients with traditional feature parameters. The experimental results show that the identification effect of MFCC1was a little better than that of MFCC. Moreover, either MFCC or MFCC1being combined with feature parameter E, it did certain help to the identification effect. Hence we came into a conclusion that the method succeeded to represent the non-stationary properties of cough sound preferably after being combined with energy.Secondly, this paper applied Empirical Mode Decomposition(EMD) to divide the signal into several Intrinsic Module Function(IMF), after that, wiped out the IMF with slow-varying tendency and directly reconstituted the original signal and reestablished the primary signal normalized weighed combination of Mel scale curve, at last extracted MFCC and MFCC1of the reconstituted signals respectively. And then combined respectively with energy E, completed the combination that similar to the recognition effect of MFCC, MFCC1combinations was good.Moreover, based on analysis and comparison of Hilbert Marginal Spectrum and Fourier Spectrum, we replaced Fourier Spectrum with Hilbert Marginal Spectrum to extract Sub-band Energy Cepstrum Coefficients(SECC), which produced the minimum rate of error in the experiment.Finally, this paper adopted Teager energy operator in the Hilbert Marginal Spectrum and extracted Sub-band Energy Cepstrum Coefficients(STECC). The experimental results show that the recognition effect of STECC is a best effect relatively.This paper presents the feature parameter STECC by combining Hilbert Marginal Spectrum with Teager energy operator. The parameter, which has a best effect relatively with97.69%of recognition rate and1.15%of error rate, provides certain reference for detecting and recognizing cough sound objectively.