The Study Of Cough Signal Recognition Based On HMM-ANN Hybrid Model

Cough is the most common symptom of respiratory disease, its parameters such as frequency, intensity, types, duration provide important information for clinical diagnosis. At the current stage, the cough evaluation is usually based on the complaints of the patients, lack of objective measurement and quantitatively evaluation standards and analysis system. With the extensive application of speech recognition and artificial intelligence, human-machine interaction is urgently hoped to be used in cough signal analysis and assessment, let the machine understand cough signal, identify cough signal and then research and analyse the cough signal.Referring to speech recognition technology and cough signal research, a new kind of HMM-ANN hybrid model is applied to cough signal recognition by analyzing HMM and ANN. Simulation experiment is conducted in MATLAB. The main works are as follows:1.The cough signal is pretreated based on the analysis of the cough signal mechanism and the acoustic characteristics. According to the characteristics of cough signal, in-depth research is conducted in the whole process, such as sampling, filtering, pre-emphasis, enframe and add window, endpoint detection etc.2.The linear forecast coefficient, linear forecast cepstrum coefficients (LPCC) and Mel frequency cepstrum coefficients (MFCC) are analyzed. Through the comparative tests, the MFCC based on auditory characteristics recognition in cough signal is more excellent than LPCC based on vocal tract model. In order to reflect the dynamic characteristic of cough signal and suppress the influence of noise, MFCC's second feature is extracted, and combined RASTA with MFCC characteristic, improved logarithm function conversion, namely, taking the MFCC + first-order differential as cough signal characteristic parameters. Simulation results show that, compared with the other three parameters, the recognition result is more improved by MFCC + first-order with the noise suppression technology.3.ANN and HMM are combined for HMM's better time-domain modeling capability and ANN's powerful classification ability, namely, the transition cumulative probability of HMM Viterbi decoding is served as the input of ANN, the final output is the probability by ANN nonlinear mapping.The cough signal HMM-ANN hybrid model is established, the hybrid model of signal in the learning and recognition is researched. Based on the above research, the result of simulation experiment in MATLAB shows that the cough signal recognition performance is improved based on hybrid model.