Speech Recognition Based On Surface Electromyographic Signals

Speech communication is an important part of interpersonal communication.Due to the inability of individuals with congenital or acquired speech function to produce sound normally,it is urgent to establish communication with others;In some special fields such as space research,aircraft piloting,noisy factories,and private environments,sound cannot propagate normally and special technical means are also needed to assist communication.Although traditional speech recognition technology has been applied,it cannot avoid the disadvantage of being susceptible to environmental noise interference in these scenarios.Speech recognition technology based on surface electromyogram(s EMG)signals can adapt well to these scenarios.The main work of this thesis is summarized in the following three aspects:(1)A dedicated collection device has been designed to address the issue of surface electromyography signal acquisition being susceptible to noise interference such as environmental noise,motion artifacts,crosstalk,and power frequency interference.This thesis first studies the muscles related to pronunciation,identifies three muscles used for speech recognition,analyzes the characteristics of s EMG,lists the sources and composition of noise,and finally designs a high signal-to-noise ratio s EMG acquisition device,including device selection,circuit schematic design,PCB design,software design,and other aspects.The entire composition of the acquisition device is presented in the form of hardware open source,A detailed design process and component list were provided.After horizontal and vertical comparative experiments,it was verified that the system requirements were met.It was found that the hardware filtering effect was intuitive and there was a significant difference in comparison,proving the uniqueness and superiority of the design.It can be used for speech recognition experiments based on surface electromyography signals.(2)A denoising application based on wavelet transform is proposed to address the timeconsuming and technically challenging process of surface electromyography signal processing.Firstly,it was explained in detail that the denoising process based on wavelet is divided into three steps: wavelet decomposition,wavelet denoising,and wavelet reconstruction.Secondly,five processing parameters need to be selected,including(a)the type of wavelet basis function;(b)the scale;(c)the threshold selection rule;(d)the threshold rescaling method,and(e)the thresholding function.Finally,experiments were carried out on each combination,and the signal-to-noise ratio was selected as the evaluation standard.It was concluded that the wavelet function most suitable for processing surface EMG signals was the fifth order Daubechies function grade 8,and the Minimax thresholding method plus Hard thresholding was the best combination of the eight possible wavelet denoising programs.(3)In response to the limited recognition accuracy of existing speech recognition based on s EMG signals,this thesis compares multiple machine learning algorithms and neural network algorithms to achieve the highest recognition rate.First,collect the s EMG signal,extract the timedomain and frequency-domain characteristics of s EMG signal for speech recognition,and then use machine learning algorithms,including linear discriminant analysis(LDA),K-Nearest Neighbor(KNN),Support Vector Machine(SVM),Bayes,random forest(RF),binary tree,Experiments were conducted on the combination of each feature and each classification algorithm,and model parameters were optimized during iteration.Finally,in order to improve recognition accuracy,the author investigates the use of speech recognition algorithms based on Long Short Term Memory(LSTM)networks to construct efficient and accurate dynamic models.The experimental results show that in machine learning algorithms,the recognition accuracy of support vector machines is82.25%;Among the neural network algorithms,LSTM has the highest recognition accuracy,reaching 85.34 %.