Research On Silent Speech Recognition Based On PSO-SVM Algorithm

Voice plays a very important role in human daily communication,but in some cases,people can not communicate normally through voice.For example,in a highnoise environment,a way of communication that can effectively transmit information is needed.And patients with throat cancer can not communicate with others through voice,so silent voice for information transmission is needed.When a person makes a vocal action in a silent state,the muscles of the face and neck related to the sound will produce different states of activity.Surface electromyography is usually used as the research object for silent speech recognition.The main research contents are as follows:(1)The surface electromyography signals of facial levator lip muscle and cervical platysma muscle were collected when people uttered 6 single vowels and 6commonly used Chinese words in Hanyu pinyin in silent state.Wavelet method and band-stop filter are used to Denoise the signal,and the denoising effects of different methods are compared.The experimental results show that the denoising effect of wavelet soft threshold is the best.Therefore,the wavelet soft threshold denoising method is selected to process the signal.(2)An endpoint detection algorithm based on Mel cepstrum distance and subband spectral entropy is proposed.The existing endpoint detection methods of surface electromyography signals are mainly based on threshold,but the accuracy of threshold-based detection methods needs to be improved.By using a single feature for endpoint detection of surface electromyography signals,it is found that Mel cepstrum distance and subband spectral entropy have better detection results,so Mel cepstrum distance and subband spectral entropy are selected as features,combined with machine learning for endpoint detection.The experimental results show that for the endpoint detection of six single vowels,the multi-feature fusion endpoint detection method can obtain lower weighted error measurements under the condition of low signal-to-noise ratio,and has good stability and robustness.(3)The inertia weight and acceleration coefficient of the particle swarm optimization(PSO)algorithm are improved to improve the search efficiency of the algorithm.The traditional ground particle swarm optimization algorithm uses fixed inertia weight and acceleration coefficient,so it is easy to fall into the local optimal solution when dealing with complex problems.The global and local search ability of particles is balanced by linearly decreasing inertia weight strategy and nonlinear dynamic learning factor.The four classification methods of naive Bayesian,Knearest neighbor,support vector machine and PSO-SVM are compared.the results show that for the classification and recognition of surface EMG signals of six single vowels in Hanyu Pinyin,PSO-SVM algorithm shows good performance,and can get 96.7% accuracy and 95.5% recall rate.For 6 commonly used Chinese words,it can achieve 91.3% accuracy and 90.5% recall rate.